Computational anatomy as a driver of understanding structural and functional cardiac remodeling.
Gabriel Bernardino
TESI DOCTORAL UPF / 2019
THESIS SUPERVISORS Bart Bijnens, Miguel Angel González Ballester and Mathieu De Craene
Department Engineering and Information and Communication Technologies Academic Coordination Unit
The cover artwork is a part of a photographic reproduction, available in Wikimedia, of a public domain domain painting: Anatomy of the heart by Enrique Simonet (1890), currently deposited in Museo de Málaga, Málaga, Spain.
III
Acknowledgements
This has been a long wandering, which I couldn’t have endured alone. All of this has been possible thanks to my supervisors: Bart, Miguel Angel and Mathieu. They helped me by putting order in chaos (and sometimes also chaos into order). Their constant guidance, advice and help have shaped this thesis. The quality of this work reflects their quality as supervisors. A very special thanks to my colleagues in UPF, partners in coffees, merendolas, beers, aperols and wine. They have made the hard moments easier to cope with, and our scientific discussions provided a lot of ideas and insights. I have to thank the people of Philips for their hospitality and the pauses gourmandes. I would like to thank the clinical collaborators, at the CHU Caen, Hospi- tal Clínic and Maternitat. Without them this thesis would be useless. Their patience explaining the most basic of the concepts of medicine and clinical cardiology has been crucial for this thesis. I also want to thank them for their time they took from their busy schedules to understand and discuss my results. A very special thanks to Eric and Amir of the CHU Caen who hosted me during a month. Finally, I want to thank my friend and family for all the support and understanding in the difficult times.
V VI Abstract
Cardiac structural and functional remodelling, induced by altered or adverse working conditions, has been extensively reported in the litera- ture. The quantification and interpretation of such remodelling is still an ongoing research topic, and the mechanisms and lasting effects are not completely understood. A difficulty in understanding remodelling is that, even if, at cellular level, each cell acts independently, it is the organ-level aggregation that will ultimately determine the cardiac efficiency. Given that remodelling is driven by regional stimuli, it will not be homogeneous but will express as complex segmental patterns. The assessment through tra- ditional methods is challenging due to the focus on global quantification of the current clinical measurements, as well as the large heterogeneity between the responses of individuals. We present a statistical shape analysis (SSA) framework to identify the expected appearance of regional shape remodelling. We use a patho- logical and control population to build a model that identifies populational differences in shape. This model consists of a dimensionality reduction step (PCA and PLS) and a classification step (logistic regression). Given the large natural shape variability present in the ventricles, that can cre- ate a confounding effect if there is an imbalance of the demographics, we applied and tested methods to account for that variability: adding the con- founders as covariates in the classification model (adjustment) and build- ing a model that predicts shape from the confounders, and analysing the regression residuals (confounder deflation). We show that these methods are able to correct for demographic imbalances. The previous methodology was applied to two distinct cardiac magnetic resonance imaging (MRI) datasets: one with triathletes and another one with small-for-gestational age (SGA) individuals, which were compared to a control population to obtain the remodelling. SGA is a syndrome oc- curring at foetal stage where the individual has an impaired growth during foetal development. It is hypothesized that it translates to a higher car- diovascular risk during seniority. Using our framework, we were able to identify that SGA presented a more curved base in the right ventricle (RV) compared with controls, especially in smokers and overweight SGA indi- viduals. Athletes were also compared to controls in order to characterize endurance-activities-related remodelling. We identified a pattern that was comprised of the already known athlete’s remodelling: an increase of the left ventricular (LV) volume size and mass, but also an increase of RV volume localized in the outflow. The found quantification of the athletic remodelling was associated to a better cardiovascular response during a
VII maximal stress test. Finally, we explore alternatives to analyse regional shape of the RV that did rely on the point-to-point registration when the imaging modality is noisy and with low contrast, such as 3D echocardiography. We imple- mented a mesh independent volumetric parcellation of the RV in three parts: inlet, outflow and apex. The parcellation was defined on the surface using the geodesic distances to the apex, tricuspid and pulmonary valve, and was propagated to the cavity using the Laplace equation. We tested the reproducibility of the parcellation and found an acceptable mean error ( 8%) in the intraobserver test, and a higher (>14%) for the interobserver. We validated the method in a synthetic-remodelling generated dataset and found that our method was accurate for capturing circumferential dilation but less suited for longitudinal elongations.
VIII IX Resumen
El remodelado cardiaco ha sido descrito extensivamente en la litera- tura. Este remodelado, que afecta tanto la estructura como la función, es inducido por una alteración de las condiciones de trabajo del corazón. Su cuantificación e interpretación es todavía un tema en investigación y sus mecanismos y efectos no han sido todavía completamente entendidos. Una dificultad para entender dicho remodelado es que, aunque a nivel celular cada célula actúa de manera independiente, es la agregación a ni- vel de órgano la que en ultima instancia determina la eficiencia cardíaca. Dado que el remodelado esta dirigido por estímulos regionales, éste no resultará en una reacción homogénea sino en complejos patrones seg- mentales. La evaluación con los métodos tradicionales es complicada ya que las medidas clínicas se centran en cuantificación de variables globa- les. Presentamos un sistema de análisis estadístico de forma (SSA) que permite identificar patrones esperados de remodelado estructural y regio- nal. Usamos una población patológica y otra de control para construir un modelo que identifica diferencias de forma. Este modelo consiste en una reducción de dimensionalidad (PCA y PLS) seguida de una clasificación (regresión logística). Debido a la gran variabilidad estructural presente en los ventrículos, un desequilibrio en los demográficos de las poblaciones puede crear un factor de confusión, hemos aplicado y testeado dos méto- dos para ajustar por esa variabilidad: el primero añade los posibles facto- res de confusión como covariables en el modelo de clasificación (ajuste) y el segundo construye un modelo que predice la forma a partir de estos factores de confusión y analiza los residuos de regresión de dicho modelo (deflación de los factores de confusión). Mostramos que estos métodos son capaces de corregir desequilibrios en los demográficos. Aplicamos la metodología anterior a dos bases de datos de imagen por resonancia magnética: uno de triatletas y otro de individuos pequeños- para-su-edad-gestacional (SGA), a los que comparamos con una pobla- ción de controles para obtener el patrón de remodelado. SGA es un sín- drome que consiste en una restricción al crecimiento durante la etapa fetal. Se ha hipotetizado que los individuos con SGA tienen una mayor propensión a problemas cardiovascular durante la tercera edad. Usando nuestro sistema, encontramos que los SGA tienen una base del ventrícu- lo derecho (RV) más curvada, especialmente en SGA fumadores o con sobrepeso. Los atletas fueron también comparados con los controles pa- ra caracterizar el remodelado relacionado con la práctica de deportes de resistencia. Identificamos un remodelado concorde con la literatura: un in-
X cremento del ventrículo izquierdo (LV) tanto en tamaño como en masa, pero también un incremento del volumen del RV focalizado en el infun- díbulo. Asociamos la cuantificación de dicho remodelado con una mejor respuesta cardiovascular durante una prueba de ejercicio máxima. Finalmente, exploramos alternativas al análisis regional de forma del RV que no usasen el registro punto a punto para cuando la modalidad de imagen fuera ruidosa y con poco contraste, como es el caso de la ecocardiografía 3D. Implementamos un método de parcelación volumétri- ca independiente del mallado que divide el RV en 3 partes: apical, inlet y outflow. La parcelación fue definida usando las distancias geodésicas al apex, tricúspide y válvula pulmonar, y se propago a la cavidad usando la ecuación de Laplace. Probamos la reproducibilidad de la parcelación y encontramos un error medio intraobservador aceptable (8 %), pero mas error (>14 %) en el caso del interobservador. También validamos el méto- do usando una base de datos de remodelado generado sintéticamente y encontramos que nuestro método era exacto para analizar crecimientos circunferenciales, pero no para elongaciones longitudinales.
XI
The only attitude worthy of a superior man is to persist in an activity he recognizes is useless, to observe a discipline he knows is sterile, and to apply certain norms of philosophical and metaphysical thought that he considers utterly inconsequential.
Fernando Pessoa, Livro do Desassossego.
XIII
Contents
List of figures XX
List of tables XXII
1. INTRODUCTION 1 1.1. Motivation ...... 1 1.2. Cardiac remodelling ...... 1 1.2.1. Morphological remodelling ...... 2 1.2.2. Functional remodelling ...... 3 1.3. Medical imaging ...... 5 1.4. Computational anatomy ...... 6 1.5. Thesis structure ...... 9
2. HANDLING CONFOUNDING VARIABLES IN STATISTICAL SHAPE ANALYSIS - APPLICATION TO CARDIAC REMODELLING 11 2.1. Introduction ...... 11 2.2. Methodology ...... 14 2.2.1. Atlas construction ...... 14 2.2.2. Confounding deflation ...... 15 2.2.3. Classification ...... 18 2.3. Experimental setup ...... 19 2.3.1. Clinical dataset ...... 19 2.3.2. Automatic measurements ...... 20 2.3.3. BMI-based downsampling ...... 21 2.4. Results ...... 22 2.4.1. Dimensionality reduction ...... 22 2.4.2. Athletic model ...... 23 2.4.3. BMI model ...... 26 2.4.4. Confounding adjustment ...... 28 2.4.5. Confounding deflation ...... 29 2.5. Discussion ...... 33
XV 2.6. Conclusion ...... 35
3. VOLUMETRIC PARCELLATION OF THE RIGHT VENTRICLE FOR REGIONAL GEOMETRIC AND FUNCTIONAL ASSESSMENT 37 3.1. Introduction ...... 37 3.2. Methodology ...... 40 3.2.1. Data acquisition ...... 40 3.2.2. Parcellation of the right ventricle ...... 40 3.2.3. Local and global anatomic frame of reference . . . . 43 3.2.4. Strain as a value to express local deformation . . . . 45 3.2.5. Synthetic regional remodeling patterns generation . 45 3.2.6. Global remodelling ...... 49 3.3. Validation ...... 49 3.3.1. Reproducibility of the 3D models ...... 49 3.3.2. Reproducibility of the parcellation method ...... 50 3.3.3. Validation of the parcellation method ...... 50 3.4. Results ...... 51 3.4.1. Reproducibility of the 3D models ...... 51 3.4.2. Reproducibility of the parcellation method ...... 55 3.4.3. Validation of the parcellation method ...... 56 3.5. Discussion ...... 59 3.6. Conclusion ...... 60
4. THREE-DIMENSIONAL REGIONAL BI-VENTRICULAR SHAPE REMODELLING IS ASSOCIATED WITH EXERCISE CAPACITY IN ENDURANCE ATHLETES 61 4.1. Introduction ...... 61 4.2. Methods ...... 62 4.2.1. Population ...... 62 4.2.2. Echocardiographic measurements ...... 63 4.2.3. Exercise Test ...... 63 4.2.4. MRI study ...... 64 4.2.5. Statistical shape analysis ...... 64 4.2.6. Statistical methods ...... 65 4.3. Results ...... 65 4.3.1. Population characteristics ...... 65 4.3.2. Athletic shape remodelling ...... 66 4.3.3. Exercise response ...... 71 4.3.4. Validation ...... 72 4.4. Discussion ...... 73 4.4.1. Limitations ...... 75 4.5. Conclusion ...... 75
XVI 5. REDUCED EXERCISE CAPACITY AND EXAGGERATED IMPACT OF CONVENTIONAL RISK FACTORS ON CARDIAC FUNCTION IN ADULTS BORN SMALL-FOR-GESTATIONAL AGE 77 5.1. Introduction ...... 77 5.2. Methods ...... 79 5.2.1. Study design ...... 79 5.2.2. Echocardiography ...... 79 5.2.3. Cardiovascular magnetic resonance ...... 80 5.2.4. Ventricular shapes ...... 81 5.2.5. Cardiopulmonary exercise testing ...... 81 5.2.6. Sample size calculation ...... 81 5.2.7. Statistical analysis ...... 82 5.3. Results ...... 83 5.3.1. Perinatal data and characteristics at recruitment . . 83 5.3.2. Baseline cardiac structure and function ...... 85 5.3.3. Exercise capacity ...... 90 5.3.4. Effect of smoking and overweight on cardiac remod- elling among the study groups ...... 90 5.4. Discussion ...... 92 5.5. Conclusion ...... 98
6. CONCLUSION 99 6.1. Summary ...... 99 6.2. Methodological future work ...... 100 6.3. Clinical future work ...... 101
XVII
List of Figures
1.1. Concentric and eccentric remodelling ...... 3 1.2. Stroke volume, cardiac output and LV dimensions during ex- ercise...... 4 1.3. Number of procedures in the different modalities at Chil- dren’s Hospital Boston...... 5 1.4. Segmented biventricular 3D model personalised from MRI .6 1.5. The heart as seen by different modalities...... 7
2.1. Schema of the SSA framework ...... 14 2.2. Comparison of different DR methods and confounding ad- justment...... 24 2.3. Most discriminating shape patterns ...... 25 2.4. Effect of adjustment in the measurement response . . . . . 25 2.5. Close-up on the right ventricular free wall of the different models predicting the athletic remodelling shape pattern . . 26 2.6. Representative shapes of the BMI model ...... 27 2.7. Measurement response of the BMI-related shape changes. 28 2.8. Stability of adjustment in BMI-imbalanced datasets. . . . . 29 2.9. Measurement response of the models trained on the down- sampled population ...... 30 2.10.Representative shapes of models trained on the downsam- pled population ...... 31 2.11.Stability analysis of the confounding deflation method trained with a non-downsampled population...... 32 2.12.Measurements response of the shape pattern found with the athlete-downsampled data...... 32 2.13.Effect of the population used to train the confounder defla- tion model on the discriminative pattern stability...... 33 2.14.Distribution of the dummy variable and its dot product with the athletic remodelling shape pattern ...... 33
3.1. Steps to generate the volumetric partition...... 42
XIX 3.2. Circumferential and longitudinal directions defined in each triangle of a sample RV mesh ...... 44 3.3. Local frames of reference and dihedral angles...... 46 3.4. Mean point-to-point registration error for each node on the interobserver and intraobserver reproducibility test...... 52 3.5. Mean point-to-surface distance for each node on the inter- observer and intraobserver reproducibility test...... 53 3.6. The two generated 3D models on the test/retest experiment and their parcellations...... 54 3.7. Generated meshes with the local remodelling method . . . 57 3.8. Generated RV meshes for the global remodelling...... 58 3.9. Volume response to a global remodelling...... 58
4.1. Sample MRI segmentation of an athlete...... 68 4.2. Sample MRI segmentation of a control...... 69 4.3. Most discriminant shape mode that distinguishes the RV from athletes and controls...... 70 4.4. LOO-CV ROC of the model which considers shape and con- founders, compared to a simple model that only considers the confounders...... 72 4.5. Quantification of the athletic remodelling score for both male and female athletes...... 73
5.1. Baseline cardiac shape of the study populations (non-stressed) 86 5.2. Workload and peak VO2 and its relationship with LV mass and remodelling score in the study populations ...... 91 5.3. Left and right ventricular 3D meshes generated by CMR from control and SGA datasets among overweight and smok- ing populations respectively...... 93 5.4. Illustrative CMR examples in an over weighted and smoker control and SGA individuals...... 94 5.5. Regression lines showing an inverse relationship between the performance in the exercise test and ventricular remod- elling score in smokers and overweight subpopulations. . . 94
XX List of Tables
2.1. Population demographics of the study participants. Athletes have a lower heart rate and weight than controls ...... 21 2.2. Results of DR model selection ...... 23 2.3. 10-fold CV log-loss results of different dimensionality reduc- tion (DR) methods parameters, ...... 23
3.1. Intra and inter-observer variability of the segmentations and node positions...... 51 3.2. Intra- and inter-observer variability of the segmental and to- tal end-diastolic volumes and EF ...... 55 3.3. Regional volumes resulting from two consecutive acquisi- tions of the same patient...... 55 3.4. Volume response to a global remodelling ...... 56
4.1. Demographics and echocardiographic functional measure- ments of the population...... 66 4.2. Comparison of the MRI- based measurements between ath- letes and controls...... 67 4.3. Values at rest, after exercise and percent ratio of function and geometry echocardiographic and ergospirometric mea- surements in athletes...... 67 4.4. MVR results between the shape remodelling score and dif- ferent functional parameters during exercise...... 71 4.5. Results of the linear regression model predicting peak VO2 and the logistic model predicting athletes...... 73 4.6. Correlation coefficient between the different classical indices used to quantify athletic morphological remodelling and our remodelling score...... 74
5.1. Perinatal and current baseline characteristics of the study population ...... 84
XXI 5.2. Perinatal characteristics and current laboratory results of the study population...... 85 5.3. Echocardiographic results of the left heart study populations 87 5.4. Echocardiographic results of the left heart study populations 88 5.5. Cardiovascular magnetic resonance results of the study pop- ulations...... 89 5.6. Results at peak exercise in the study populations...... 90
XXII Acronyms
AUC area under curve
BMI body mass index
BSA body surface area
CO cardiac output
CV cross validation
DR dimensionality reduction
ED end diastolic
EF ejection fraction
ES end systolic
HCM hypertrophic cardiomyopaty
HR heart rate
IUGR intrauterine growth restriction
LH left heart
LR logistic regression
LV left ventricle
LVGLS left ventricular global longitudinal strain
MAPSE mitral annular systolic excurse
MRI magnetic resonance imaging
MVR multivariate regression
PCA principal component analysis
PDM point distribution model
PER pulmonary extraction rate
PLS partial least squares
RA right atria
XXIII RH right heart
ROC receiver operating characteristic
RV right ventricle
RVOT right ventricular outflow tract
SA short axis
SGA small-for-gestational age
SSA statistical shape analysis
SV stroke volume
TAPSE tricuspid annular systolic excurse
VCO2 carbon dioxide production
VO2 oxygen uptake
XXIV Chapter 1
INTRODUCTION
1.1. Motivation
The aim of this thesis is to develop and apply computational tools to analyse, understand and quantify cardiac structural remodelling patterns. The heart experiences changes to adjust to altered loading conditions in both a chronic and acute manner. The adaptations occur to maintain a certain number of (functional) parameters within a physiological range, but, while beneficial at some point, sometimes they have side effects that can become adverse in the long-term. The adaptation mechanisms show large heterogeneity: different individuals may remodel differently to the same stimulus. Understanding and quantifying the remodelling and its effects is relevant for diagnosis and prognosis, allowing to identify individ- uals at risk of having adverse remodelling. Despite the success of com- putational anatomy in the biomedical engineering community, analysis of remodelling in the clinic relies heavily on single-measurement based tech- niques, that are ill-suited to discriminate more subtle or complex regional patterns. In this thesis, we develop and use computational anatomy tech- niques to study subtle regional cardiac morphological remodelling and its association to functional changes.
1.2. Cardiac remodelling
As stated above, the heart is not the same through a whole life and has many compensatory mechanisms to respond to changes in the heart/body. These changes can be acute, like increasing heart rate to increase blood delivery during exercise, or can be chronic, like the thickening of the my- ocardium in the left ventricle (LV) of hypertensive patients. Overall, these
1 mechanisms try to maintain a sufficient cardiac output (CO) to provide enough oxygen to the system, but also maintain the wall stress low so the cardiac myocytes don’t get damaged and die.
1.2.1. Morphological remodelling
In the past, morphological/structural remodelling has been simplified as a dilatation (or eccentric) remodelling reaction as a result of volume overload, and an increase of myocardial mass (concentric) induced by pressure overload (Grossman et al., 1975). However, this oversimplifi- cation, often referring to the whole ventricle, does not match reality: the remodelling process at cellular level is a complex mechanism where each myocyte acts independently to the others, based on the local signalling re- ceived (Opie et al., 2006). The different cellular remodelling pathways are triggered not only by mechanical stress endured by the myocyte, but also by the presence of, for example, hormones. Hypertrophic growth of the cell has been associated to systolic stress, but myocyte elongation is more complex and is still under discussion. Therefore, the resulting remodelling is not homogeneous in the whole heart, but overall expresses as aggrega- tion of all the individual cellular changes. These regional patterns are of diagnostic interest since they can allow to determine the cause of the re- modelling, by inspecting the differences in remodelling magnitude among different areas, and the type of remodelling present. Even if the cellular reaction is depending only on the local stimulus, the relationship between ventricular pressure and wall stress is dependent of the local cardiac shape. Laplace’s Law describes the relationship between pressure and stress for thin-walled surfaces (Basford, 2002). The gener- alised Young-Laplace equation, where K notes the curvature tensor, the stress tensor, h the wall thickness and Tr the trace operator is:
P = T r(K)/(2h) (1.1)
As stated by Laplace’s Law, for the same pressure, wall stress de- creases for more curved walls but increases for flatter ones, and thicker walls are able to withstand pressure with lower stress. Therefore, remod- elling to lower stress due to a pressure increase is a remodelling that does not only make the wall thicker but also goes towards more globular hearts (Ganau et al., 1992; Zhong et al., 2009). While differences in remodelling produce shape changes, differences in shape also trigger remodelling: the base of the LV is the flattest, and therefore the one that undergoes bigger stresses when pressure is homo- geneous, inducing potentially more remodelling (Baltaeva et al., 2007).
2 Figure 1.1: Concentric and eccentric remodelling. Reproduced from (Gjes- dal et al., 2011)
1.2.2. Functional remodelling
Not only the cardiac shape can change, but also the function can adapt. For instance, a common observation is the variation of heart rate (HR) and blood pressure to match alterations of systemic oxygen needs throughout the day. The capacity of the heart to acutely adapt to different situations is very important, and an indicator of early stages of disease: a patient can be asymptomatic at rest, but the heart is unable to adapt to changes. Usu- ally, this lack of adaptation capability will worsen over time, leading to heart failure. A good example is exercise: during exercise, the need of oxygen increases (Poole et al., 1997). How the heart acutely responds to exercise is very important and assessing this response can give information of the overall state of the heart (Sitges et al., 2017). An important mechanism of acute remodelling is the Frank-Starling law, which states that that contraction force of the cardiac myocites increases when the end diastolic (ED) volume increases (Katz, 2007). The increase of volume stretches the myosin in the sarcomeres and allows more of their heads to attach to the actin, resulting in more force development. However, when the sarcomeres are too stretched, the myosin heads separate from the actin and there is a loss of contractile force. This mechanism is very
3 Figure 1.2: Stroke volume, cardiac output and LV dimensions during exer- cise for controls and patients who underwent Fontan procedure. We can observe that, in controls, during maximal exercise, LV volume starts de- creasing (due to lack of filling) and so does SV. Data from Claessen et al., 2019
important during light/moderate exercise, because it increases ejected vol- ume as a reaction to the increased venous return. When the exercise increases in intensity, the heart increases HR, thus shortening diastole. When diastole is too short, then stroke volume (SV) starts to decrease because the heart is not getting enough blood to fill.
It is important to understand that geometry and function are not inde- pendent, but intimately related. Changes in function affect geometry, and geometry influences function. For instance, endurance athletes present dilated hearts to increase their SV. Since their CO demand at rest is main- tained, they lower their resting HR, and often also muscle deformation, because they can. Understanding this interplay is crucial for assessment of the remodelling in order to determine whether it is adaptative or mal- adaptative.
4 Figure 1.3: Number of procedures in the different modalities at Children’s Hospital Boston. Echocardiography shows a steady increase and is cur- rently the most common procedure. Data from Prakash et al., 2010
1.3. Medical imaging
In cardiology, the workhorse of imaging is echocardiography (>95% of the cardiac images generated are using this modality, Figure 1.3). 2D echocardiography allows assessment of both function and geometry. It is cheap (compared to other modalities), can be operated at the bedside and does not use ionising radiation. It has a good temporal resolution and short scanning times. Function can be assessed using B-mode and Doppler imaging to obtain blood or tissue velocities, or using speckle tracking on the walls to obtain strain. There are also derived indices, like ejection fraction (EF) or tricuspid annular systolic excurse (TAPSE)/mitral annular systolic excurse (MAPSE), that assess function by looking at the deformation of cardiac walls. The geometry of the LV can be reasonably assessed with simplified measurements, due to its symmetry, thicker walls and closer position to the chest. However, the RV has thin walls, is more trabeculated and cannot be easily imaged with 2D echocardiography due to its position in the chest (Figure 1.4). The RV asymmetric geometry also makes 2D- based assessment difficult, as the use of assumptions in the geometry introduces distortion. However, 2D echocardiography has many shortcomings: it is very de- pendent on the experience of the operator, and some patients have poor echocardiographic windows that produce images full of artefacts. Since it is 2D and the probe is positioned manually, it is difficult to acquire exactly the same images each time, resulting in a relatively low reproducibility of measurements (D’hooge et al., 2016). It has a low contrast to noise ra- tio, that increments with the distance from the transducer, hampering atrial assessment when viewed from the apical window. 3D technologies offer
5 Figure 1.4: Biventricular 3D model personalised from MRI. The RV is coloured in blue and the LV is coloured in red. We can see that while the LV resembles a prolate ellipsoid, the RV has a more triangular and biaxal shape. advantages, but at the expense of higher costs and lower temporal, as well as spatial, resolution. In this thesis, we have used 3D data acquired with MRI and 3D echocar- diography. 3D techniques allow to improve reproducibility (Thavendiranathan et al., 2012), since they are less dependent on the acquisition plane, but this comes at a cost. Acquisition is longer: a 3D volume of the heart acquired by MRI requires several apneas, and 3D echo can be usually acquired in a single apnea of 4-6 cardiac cycles. This longer acquisition results in the need of using ECG gating and image compounding, and the generation of artefacts from the fusion of subvolumes. MRI cannot be op- erated at the bedside but needs the patient to be transported to the device. 3D echo has lower image quality than 2D echo. MRI has a relatively large pixel size but has a better contrast that allows for more easy automatic segmentation of the structures.
1.4. Computational anatomy
3D shapes are difficult to analyse by humans due to their high complex- ity and large data volume. As stated above, remodelling results in complex regional patterns that are difficult to express in a simplified quantitative manner. The use of computational techniques for its analysis allows to, not only quantify images, but also understanding them. These techniques allow to analyse data that is too big and tedious to be analysed by humans alone. Figure 1.4 shows an example of the shapes that are analysed in this thesis: the cardiac ventricles obtained from magnetic resonance imaging (MRI). We can see that the LV presents a regular, convex and symmetric shape. On the other hand the right ventricle (RV) resembles no simple
6 (a) B-mode echocardiography (b) Color Doppler echocardiography
(c) MRI
Figure 1.5: The heart as seen by different modalities. geometric, thus requiring more sophisticated computational techniques for its analysis. Computational anatomy is the discipline that studies quanti- tatively the shape of organs. It intersects with other disciplines, such as medicine, mathematics and computer vision. Statistical shape analysis (SSA) is a subfield of computational anatomy that analyses organ shape variability(Dryden et al., 1998), learning it from a population(Cootes et al., 1995; Joshi et al., 2004). It has been exten- sively applied: as segmentation (O’Brien et al., 2009; Heimann et al., 2009), to recover complete organs from partial observations (Albrecht et al., 2013), or to compare populations (Zhang, Cowan, Bluemke, Finn, Fon- seca, Kadish, Lee, J. A. Lima, et al., 2014; Sarvari et al., 2017). In the latest years, SSA has been combined with machine learning techniques to build models that outperform, at least in the population used for training, traditional methods (Bernard et al., 2018; Gilbert et al., 2019; Bhuva et al., 2019). SSA allows to operate in the space of shapes directly, without being restricted to a few pre-specified measurements. It allows to identify and quantify complex shape patterns that would be otherwise difficult to find by simple observation. It is therefore a very potent tool to find remod- elling patterns in an exploratory way, without imposing a hard constraint on which kind of patterns the researcher is trying to find.
7 SSA is a powerful tool, but it has shortcomings. The quality of the re- sults often relies heavily on the quality of the point-to-point registration, which is an expensive and unstable process. The heart presents high shape variability and not many common anatomical landmarks that can be identified in the images. Therefore, point-to-point correspondence cannot be verified and registration uses pseudo-landmark defined through math- ematical properties. Moreover, since it learns from populations, it is very dependent of the dataset that was used to construct the models. SSA can propagate biases existing in the datasets, which is very undesirable. Fi- nally, organ shapes are elements with a very high dimensionality, and most medical image datasets involve only sample sizes of the order of hundreds. There is a big risk of overfitting the model to the sample, instead of obtain- ing generalisable results. For the cases were SSA is not applicable, due to bad registration, or very few available samples, it is possible to generate shape features like volumes, curvatures or diameters. These measurements do not need to be only global since (meaningful) parcellations of the organ can be used. Knowledge on physiology and anatomy can be imposed to define problem- specific features based on educated guesses of what information is the most relevant. While this approach has been overlooked in recent years as opposed to more organ-agnostic approaches, like SSA, it is still useful and results in more interpretable models. It can also complement SSA, using SSA to identify and understand the remodelling, and developing more sta- ble techniques to obtain remodelling image biomarkers that capture spe- cific patterns.
8 1.5. Thesis structure
In this thesis, we present novel methods to identify and assess remod- elling of the cardiac chambers in different, clinically challenging, settings. The applicability of our methods is in clinical research and diagnosis, and therefore we need them to be easily interpretable and reliable models. We have avoided addressing the steps of image acquisition and reconstruc- tion, as well as most of the process of segmentation and 3D mesh recon- struction from medical images, where we used previously developed meth- ods. The chapters (except the introduction and conclusion) are mostly self-contained and presented in the form of papers for academic journals.
2 In Chapter 2 we present a statistical shape analysis framework to robustly identify shape differences between two populations and val- idate it against demographic imbalances in the populations.
3 In Chapter 3 we use the previous framework to identify shape dif- ferences between endurance athletes and controls and link these resting shape differences to the cardiac performance in a maximal stress test.
4 In Chapter 4, we apply the SSA framework to analyse an adult popu- lation who had growth problems during their fetal gestation, compar- ing them to controls.
5 In Chapter 5, we propose a novel method for volumetric parcellation of the right ventricle in 3 regions (inlet, outflow and apex) based on geometric properties of the surface. This method allows to avoid the registration step and can work with small populations without enough samples to use the previous statistical techniques.
6 In the conclusion, we summarise the contributions and propose fu- ture directions of research to continue the lines of research present in this thesis.
9
Chapter 2
HANDLING CONFOUNDING VARIABLES IN STATISTICAL SHAPE ANALYSIS - APPLICATION TO CARDIAC REMODELLING
2.1. Introduction
Analysing the shapes of parts of biological organs and organisms has been the object of extensive study for over a century (Thompson, 1942). This interest in shape is also present in medicine: several studies have focused in the relationship between organ morphology and illness. For in- stance, cardiac shape remodels to improve cardiac pressure/volume out- put under abnormal working conditions, and it is used to assess the pres- ence/evolution of illness (Arts et al., 1994; Grossman et al., 1975). In a nutshell, pressure overload produces concentric remodelling (thickening of the myocardium without dilation of the ventricle) to maintain wall stresses low, and volume overload dilates the ventricle without a myocardial mass thickening. This is an oversimplication, as a volume overload will also in- crease pressure, and the exact remodelling mechanisms and triggers at a cellular level are still under research and discussion. The classical way
This chapter is adapted from: Bernardino G, Benkarin O, Sanz de la Garza M., Prat-Gonzàlez S., Sepulveda A., Crispi F., Butakoff C., de Craene M., Sitges M., Bijens B., González Ballester M.A. "Handling confounding variables in statistical shape analysis - application to cardiac remodelling" [Under review]
11 of analysing shape in the clinical community consists in manually extract- ing hand-crafted features, and analysing these shape descriptors. These measurements are usually standardised and defined in guidelines (Lang et al., 2015), and usually refer to global characteristics of the shape that carry little regional information. Nowadays, it is possible to acquire 3D images in clinical routine. Fur- thermore, advances in computing permit to automatically segment the im- ages and to generate personalised 3D models of the organs (González Ballester et al., 2000; Mitchell et al., 2002; Ecabert et al., 2006; Bernard et al., 2018). statistical shape analysis (SSA) is a set of techniques to rep- resent both shapes and images and do the analysis directly with these ob- jects, not being limited to only analyse previously defined measurements. SSA is used in the medical imaging field, in order to identify and repre- sent shape variability of those organs (Cerrolaza et al., 2015; Blanc et al., 2012; Rajamani et al., 2007; Sierra et al., 2006). This allows expressing and quantifying regional shape patterns in a robust and objective man- ner, instead of working with a small set of predefined measurements on the shapes (like volumes and diameters). SSA can be used together with statistical learning techniques to construct models that find regional dif- ferences in anatomy that are associated to pathologies (Zhang, Cowan, Bluemke, Finn, Fonseca, Kadish, Lee, J. A. Lima, et al., 2014; Singh et al., 2014; Varano, Gabriele, et al., 2017; Sarvari et al., 2017), based on a con- trol and a pathological population. Roughly, the typical framework consists of first using SSA to construct an atlas of all shapes in the population, then use principal component analysis (PCA) or another dimensionality reduc- tion (DR) technique to find a low-rank representation of the shape space, and finally use a classification algorithm on that space to train a model that predicts the control/pathologic status. Beyond pathological variability, the shape of an organ also exhibits vari- ability due to other factors, like lifestyle, gender, ethnicity, or size. The framework described above uses the implicit hypothesis that differences in shape are only due to the pathology. In some cases the remodelling as- sociated to the pathology is prominent and easily identifiable. However, in others, like subclinical or early-stage studies, differences can often be very subtle and less pronounced than demographic-related variability. In the lat- ter case, an imbalance in the population demographics may result in wrong associations between shape differences caused by demographics and the illness. Even in cases where the pathological remodelling is significant, and the populations are similar in terms of demographics, demographic- related variability will add noise to the analysis.A number of authors have explored the usage of non-imaging information. For instance Singh et al.
12 proposed a procedure similar to partial correlation between shape and several clinical variables while correcting for confounding variables(Singh et al., 2014). Zhang et al adjusted by demographics in their studies of car- diac remodelling in myocardial infarction (Zhang, Cowan, Bluemke, Finn, Fonseca, Kadish, Lee, J. A. C. Lima, et al., 2014), and Zhang et al and Mauger at al explored the relationship between shape and classical clini- cal measurements (Zhang, Medrano-Gracia, et al., 2017; Zhang, Cowan, Bluemke, Finn, Fonseca, Kadish, Lee, J. A. C. Lima, et al., 2014; Mauger et al., 2019). However, not all authors include corrections for confounders, and their effect in shape analysis studies has not been yet quantitatively tested. In this chapter, we present a SSA framework to find differences be- tween control and pathological populations that outputs the most discrim- inating shape pattern, that can be visualised for interpretability. We quan- titatively and qualitatively show the effect an imbalance in the confounding variables has in the analysis, and propose techniques to reduce that effect. The proposed model consists of the following steps: (1) the construction of an atlas of the personalised 3D meshes automatically generated from im- ages; (2) identification and removal of shape variability due to confounding variables; (3) dimensionality reduction and classification. To illustrate the framework, we use a dataset of cardiac magnetic res- onance imaging (MRI) involving sedentary controls and triathlon athletes. This dataset was collected to study the remodelling due to the extended practice of endurance sport, which produces a volume overload to the heart. This volume overload triggers compensatory mechanisms to im- prove cardiac output and withstand the increased pressure during exer- cise. The whole of this remodelling is called the Athlete’s Heart and involves substantial changes in function and geometry at both rest and during exercise (D’Andrea et al., 2015; Schiros et al., 2013). Although the remodelling is not yet completely understood, researchers have estab- lished a strong relationship between cardiopulmonary performance during exercise and cardiac geometry at rest (La Gerche, Burns, et al., 2012; Scharhag et al., 2002).
13 2.2. Methodology
The full process to compute the confounder invariant most discriminat- ing shape pattern between two populations is summarised in Figure 2.1. The presented framework consists of 3 main steps: 1. Compute the mean shape of the population, and register all shapes to this template.
2. Identify and remove shape variability attributable to confounding vari- ables.
3. Train a classification model to obtain the most discriminating shape pattern between both populations and generate a visual representa- tion of the most discriminating shape pattern
Procrustes alignment Original images 3D reconstruction Text
Most discriminating shape pattern
Confounding Demographics deflation Classification Dimensionality reduction
Figure 2.1: Schema of the framework and its different components. The input are the short axis MRI, and the demographics of the population. From the image, we generate a personalised 3D mesh of the ventri- cles, and align them using Procrustes analysis. Afterwards, we remove confounding-related shape variability using confounding deflation. This is followed by a dimensionality reduction and a classification steps. The final step is to compute the most discriminative shape pattern from the classifi- cation model coefficients.
2.2.1. Atlas construction
From the short axis (SA) MRI sequence, we use a model-based au- tomatic segmentation method to obtain personalised meshes with point- to-point correspondence (Ecabert et al., 2006). The method deforms a
14 full heart (4 chambers) template mesh using a polyaffine deformation to match the myocardial boundaries. The method includes slice correction to remove misalignment between consecutive slices. Since only the ventricles were visible in our images, we discarded the atria and big vessels from each segmented shape. Each resulting mesh has 4446 vertices (the left ventricle (LV) has 3052 vertices and the right ventricle (RV) has 1776 vertices, the right-most part of the septum be- longs to both ventricles) in point-to-point correspondence, and 9004 trian- gles. Only the end diastolic (ED) frame is selected for the analysis. Since meshes are in point-to-point correspondence and share the same con- nectivity, they can be analysed using the point distribution model (PDM) (Cootes et al., 1995). In PDM, the shape of each patient j is associated to a vector with the concatenated position (x, y, z) of its nodes, giving a shape vector:
Xj = (x0, y0, z0, x1, y1, z1 . . . xN , yN , zN ) (2.1)
We chose PDM over other possible representations of shape due to its simplicity and ease of computation. We applied generalised Partial Pro- crustes Analysis (Dryden et al., 1998) to align the meshes thus removing the positioning and orientation variability. We maintained size during this step, but differences in size due to anthropometric variables will be iden- tified and removed in other steps of the framework. This algorithm is an iterative method, at each iteration computes an estimation of the mean shape X¯ j and then rigidly registers each shape to that estimated mean. This is repeated until convergence, in order to obtain an unbiased mean. Here we show a full iteration j of the algorithm:
n 1 X X¯ j = Xj (2.2) n i
j j ¯ j 2 ∀i : Ri , ti = arg min kR(Xi − t) − X k (2.3) R,t
j+1 j Xi = Ri (Xi − ti) (2.4)
2.2.2. Confounding deflation
To identify and remove the shape variability related to the confound- ing variables (M), and not to the studied condition, we use a procedure similar to partial correlation, as done in (Singh et al., 2014) for a regres- sion problem. We assume that shape X can be decomposed in the sum of the population mean (µX ) and some deformation from that mean that
15 is composed of noise (), variability caused by the confounding variables (XM ) and variability from other sources (Xi). The last component includes remodelling due to pathology :
X = µX + XM + Xi + (2.5)
The first step consists in estimating XM by building a linear model (whose coefficients are wM ) that predicts the expected shape from the confounding variables M. The training data of this model is only one pop- ulation (the controls), to avoid introducing possible inter-population differ- ences in the model. This prediction will be our XM . Then, for the shape of each individual, we subtract the predicted shape from the actual shape ob- taining the prediction residuals. Residuals represent the part of the shape variability that cannot be explained by the confounding variables. To main- tain the residual vectors in the same range as the original shape vectors, we summate the original population shape mean. The final formula reads:
Xres = X − XM = (X − wc · M) + µX (2.6)
The regression model coefficients wc are computed using partial least squares (PLS) with the shapes X and confounding variables M. PLS is a regression method that projects the input and output data to two low- dimension subspaces (called embeddings) that have maximal covariance (Wegelin, 2000). The confounding variables are standardised to have 1 standard deviation and 0 mean, but shapes are only standardised to have 0 mean. The full process is described below, and consists in an iterative process where at each iteration a new regression dimension of the low- dimensional spaces is computed. The predicted part is then removed from the input and output spaces, and the process is repeated until the desired number of iterations is reached. Several versions of PLS exist, and we used Wold’s version. All PLS versions agree in the first iteration, but give different result for embedding spaces of more than one dimension. Here we show the algorithm for the r-th iteration of a PLS associating an input space X with some response Y , both being matrices. In our setting, and only for this part X will be the confounding variables(M in the rest of the chapter) and Y the shape vectors(X in the rest of the chapter). First, we compute the new dimensions of the embedding spaces by solving an eigenvector problem:
ur, vr = arg max u∗tXtY v∗ (2.7) ku∗k=1,kv∗k=1
We compute the rank one approximations (hat variables), and the re-
16 gression coefficient wr; for OLS refers to the classical ordinary least squares.
r r r r r Xˆ = OLSprediction(X u ,Y v ) (Xrur)tY rvr wr = vr(ur)t (2.8) Xrur, r r r r r Yˆ = OLSprediction(Y u ,X v )
These rank one optimisations are used to update the input X and its output Y by removing the part that is already predicted.
Xr+1 = Xr − Xˆ r (2.9) Y r+1 = Y r − Yˆ r
The final result of the algorithm is the regression coefficients w from X to Y ; and it is obtained by summing all the one-rank approximation regression coefficients. X w = wr (2.10) r
The coefficients wi of the prediction model, associated to confounder variable M i, of this model can also be visualised and interpreted. We can consider the partial determination coefficient, and visualise the shape pat- tern most associated to a certain confounder, as if it were a discriminating shape (see subsection 2.2.3).
Dimensionality reduction
Given the high dimensionality of the shape vectors and low number of samples, we use a DR method to find a subspace that contains the most relevant shape patterns. The general linear DR model reads:
Xorig = µX + KXred + (2.11)
Each shape Xorig is expressed as the population mean(µx) plus some shape-specific associated low-dimensionality vector Xred and a noise term . The embedding matrix K is constructed depending on the dimension- ality reduction method, and contains the most interesting shape directions according to a certain metric. We use three different methods of linear DR: PCA, PLS and a combi- nation of both. PCA and PLS have been reported to be used in conjunc- tion to classification methods, and in particular logistic regression (Bastien, 2005). We have described the regression modality of PLS in subsection
17 2.2.2, but PLS also computes a DR space. The difference is that we dis- card the regression coefficients w and only use embedding vectors ur in equation 2.7 of the input space. These vectors define a vector subspace, but are not guaranteed to be orthogonal, so we use QR decomposition (Lord et al., 1999) to obtain an orthonormal base. The combined method is based on prefiltering the shape using PCA, keeping a high number of components (> 90% of the variance), to then use as input to a PLS. This decreases the computation time, and adds stability by denoising the data. Contrary to the typical procedure in machine learning, we chose not to standardise the PCA modes by variance before applying PLS, as the vari- ance of each PCA mode carries important information of the signal-to- noise-ratio.
2.2.3. Classification
In this section, we present the method used to train the classifier model and use the model coefficients to obtain the most discriminating shape. The shape features obtained from the DR are combined with the confound- ing variables in a logistic regression model. We choose logistic regression because we expect not to have a complete separation between both popu- lations, and we want the model to be simple and interpretable. The logistic model gives a probability that an individual j with shape Xj and confound- ing variables Mj belongs to the pathological or control populations.
P r(yj = control |Xj,Mj) = logit(hXj, wX i + hMjwM i + b) (2.12)
Logit refers to the logistic function x 7→ 1/(1 + exp(−x)). wX and wM are the logistic regression coefficients for the shape and confounding vari- ables respectively, and are chosen to minimize the log-loss of the proba- bility of the training data X. The log-loss is the logarithm of the probability that the model is inconsistent with the observed data:
n 1 X log (|y − P r(y = 1|X ,M )|) (2.13) n j j j j
The logistic regression coefficients associated to the shape wX , can be mapped from the reduced shape space to the full space by using the pseudoinverse of the dimensionality reduction matrix. Let wXred be the coefficients associated to the reduced shape models, and KPCA and KP LS the projection matrices of PCA and PLS respectively. Then the coefficients associated with the full shape are:
t t wX = KPCAKP LSwXred (2.14)
18 Then, we can visualise and interpret the shape pattern. Analogously to multivariate regression, where the coefficients are indexed by the stan- dard deviation to allow comparison among them, we need to adjust for differences in variance of the different coordinates. Since node coordi- nates carry no meaning on their own, we treat shape as an object itself and do PCA whitening of the shape, as is typically done with other multi- dimensional signals (Kessy et al., 2018). Since we are only interested in the remodelling direction, we normalise the vector to be unitary in the L2 norm. With these corrections, all the shape features are correctly scaled by their importance in prediction. The full process to find the standardised shape pattern wˆ reads:
Σ1/2w wˆ = X (2.15) 1/2 kΣ wX k Where Σ is the covariance matrix estimated using PCA. For visualisa- tion, we can generate shapes that are representative of that shape pattern by adding the remodelling shape pattern, scaled with a parameter λ, to the mean shape. To keep the shapes within the original range, we impose that λ has to be within 3 standard deviations of the variance associated to the shape pattern.
Xrepr(λ) = µX + λwˆ (2.16) We can quantify the presence of remodelling in each shape, obtaining a scalar score for each individual, by computing the dot product of the shape vectors with the raw logistic regression coefficients associated to the shape only. The previous PCA whitening is only done for visualisation and comparison of modes. If the shape pattern needs to be quantified in a population, the original one without standardization needs to be used.
scorei = hwX ,Xi − µX i (2.17)
The shape patterns can be compared using the L2 dot product between standardised shape vectors, which coincides with the correlation of the scores associated to each pattern.
2.3. Experimental setup
2.3.1. Clinical dataset
The study comprises 77 controls and 89 athletes that underwent a MRI, to study the cardiac remodelling triggered by the practice of endurance
19 sport. The study was approved by a local ethical board, and all partic- ipants gave written informed consent for the handling of their data. Re- cruited athletes had been training an endurance sport, triathlon, over 10h a week during the last 5 years. None of the study participants had previous cardiovascular illnesses, nor any were detected during the study. Table 2.1 shows the demographics of both populations. The controls and athletes come from different studies, and the demo- graphics of both populations did not match exactly in age, but roughly rep- resent the same general population in age and gender. The study protocol and radiologist were the same for both cases. Age is statistically different between athletes and controls, but the difference is very small (2 years). We do not expect big differences due to this imbalance, since both athletes and controls are middle-aged. There are also statistically significant differ- ences in both weight and body surface area (BSA), but these correspond to physiological remodelling since endurance athletes are obviously fitter than the general population. We used as possible confounders age, BSA and gender, the typical adjustment variables in cardiology studies. The MRI acquisition was ECG-gated from the R-peak during breath- hold. The MRI machines were Siemens Aereo and Siemens Magnetom, with an in-plane spatial resolution ranging from 0.5mm to 1mm. The spac- ing between slices range between 8mm and 10.4mm, and the slice thick- ness was 8mm. MRI sequences were acquired with 25 frames per cardiac cycle. The ventricular contours (epicardium and endocardium in the case of the LV, and only epicardium of the RV ) were automatically segmented from the MRI SA using the automatic procedure described in the method- ology. The quality of the automatic segmentations was assessed by one of our experts, but no manual refinement was performed in order to preserve point-wise correspondences. Cases where errors could not be considered to be small were discarded: two individuals (both of them athletes) were discarded because the segmentation was inconsistent with the image. The segmentation and registration errors are handled as noise in our study. The meshes were very uniform and we found no self-intersecting artefacts. As a consequence of the thick slices, the apex was not correctly imaged and presented much more noise than the basal part of the ventricles.
2.3.2. Automatic measurements
We computed automatic measurements of the 3D shapes that are analogous to the classical clinical measurements, using the point-to-point correspondence and labelling coming from the model-based registration
20 Table 2.1: Population demographics of the study participants. Athletes have a lower heart rate and weight than controls. The age is significantly different, but both cohorts are middle-aged and we do not expect major age-related differences. The p-values are obtained using a Mann-Whitney test.
Athletes Controls p-value Age [y] 35.4(6.1) 33.4(3.8) 0.013 BSA [m2] 1.78(0.19) 1.86(0.20) 0.005 Weight [kg] 66.8(11.3) 73.5(15.1) 0.001 Height [m] 1.71(0.09) 1.73(0.08) 0.151 Women [%] 0.48 0.44 0.938 Resting HR [bpm] 57.2(8.4) 65.8(10.6) <0.001 and segmentation. This allowed to get a better understanding of the dis- criminative shape patterns by assessing how these measurements vary in response to the remodelling score λ on the synthetically generated meshes according to equation 2.16. We computed the ED volumes of both ventricles, as well as the myocardial mass of the LV.
2.3.3. BMI-based downsampling
Obesity (defined as body mass index (BMI) > 30) and overweight (de- fined as BMI > 25) have been reported as risk factors to cardiovascular illnesses in the literature and have a clear influence in cardiac shape and function (Alpert et al., 2018). Surprisingly, overweight and athletic remod- elling share similarities. Even if one might expect them to be opposite, as endurance athletes and overweight body fat are in the opposite sides of the spectrum, both remodellings are triggered by an increase of the heart’s loading. In the case of the athletes, it is the increase of cardiac output (CO) during exercise that produces a volume overload, and in the obese it is a mix of increased CO needs at rest to account for the bigger body size (Lavie et al., 2007) and a ventricular pressure overload due to an increase of arterial pressure (Messerli et al., 1982). To study the effect of an imbalance in a shape-affecting variable be- tween the control and case populations, we biased our population to in- crease the BMI of the controls, as overweight has a well known effect on the heart. To generate the imbalance, the control class was downsam- pled to maintain only 25% of its controls, favouring keeping the ones with higher BMI. The individuals to remove were selected randomly among the
21 controls, with a probability of being kept proportional to the rank of its BMI in the control population. Athletes were not downsampled. This proce- dure was repeated with 100 different seeds to obtain different imbalanced datasets and add robustness to the results. We analysed how this imbalance affected the found remodelling pat- tern, and to which extent could confounding adjustment and confounding deflation correct this effect. To study the stability of the shape pattern, we computed the L2 product between the discriminative shape patterns ob- tained with the downsampled datasets and the results obtained with the full dataset, that serves as groundtruth. This was performed for the dif- ferent DR methods. We also tested the L2 product of the discriminative patterns with the BMI-shape pattern, obtained with an adaptation of our framework to the regression problem. Additionally, we compared qualita- tively the automatic measurement response to the remodelling score of both the downsampled and original dataset. We tested both the covariate adjustment, and the confounding deflation. For the confounding deflation, we also evaluated how different choices during the training of the shape prediction model affected the obtained most discriminative shape pattern.
2.4. Results
2.4.1. Dimensionality reduction
To choose the best configuration of parameters for the DR method, we used 10-fold cross validation (CV) and computed the mean log-loss of all the validation set, defined in equation 2.13, over a wide combination of pa- rameters. In table 2.2 we find the log-loss of the best parameter choice for the 3 different DR methods (PCA, PLS and PCA + PLS). They correspond to a PCA with 5 modes, a PLS with 3 and the PCA + PLS with 20 PCA modes and 3 PLS modes. Table 2.3 shows the results of all the combina- tions tested. This experiment also provided an overview of how the use of demographics affected the classification metrics: there is a considerable improvement in terms of log-loss when the confounding variables are used in the model, and a minor improvement when using PLS instead of PCA. When using PCA + PLS, the metric is very similar to PLS. Confounding deflation gave a worse result than the raw (non-deflated) shapes when adjustment is used. However, confounding deflation improved the result- ing classification metric with respect to the raw when confounders are not added to the logistic regression model.
22 Table 2.2: 10-fold CV log-loss scores of the best choice for each DR method.
No deflation No deflation Method No adj. Confounders adj. No adj. Confounders adj.
PCA5 0.58 0.46 0.48 0.46 PLS3 0.59 0.44 0.52 0.48 PCA20 + PLS3 0.59 0.43 0.50 0.46
Table 2.3: 10-fold CV log-loss results of different DR methods parameters,
No deflation Confounder deflation No adj. Conf. adj. No adj. Conf. adj. Method
PCA3 0.61 0.45 0.49 0.45 PCA5 0.58 0.46 0.48 0.46 PCA10 0.62 0.49 0.52 0.48 PCA15 0.65 0.55 0.53 0.54 P LS2 0.60 0.44 0.49 0.45 P LS3 0.59 0.44 0.52 0.48 P LS4 0.66 0.48 0.51 0.50 PCA10 + P LS2 0.60 0.44 0.49 0.45 PCA10 + P LS3 0.59 0.43 0.49 0.45 PCA10 + P LS4 0.60 0.45 0.50 0.47 PCA15 + P LS2 0.60 0.44 0.48 0.45 PCA15 + P LS3 0.59 0.43 0.50 0.46 PCA15 + P LS4 0.62 0.46 0.49 0.47 PCA20 + P LS2 0.60 0.44 0.48 0.45 PCA20 + P LS3 0.59 0.43 0.50 0.46 PCA20 + P LS4 0.63 0.46 0.49 0.47
2.4.2. Athletic model
We applied our framework to identify the athletic remodelling in our dataset, and compared the effect of using the different DR method and confounding-bias correction methods (confounding deflation and adjust- ment). We used the DR parameters found in the previous section via CV. Figure 2.2 shows the L2 product between all combinations of DR methods and the adjustment or not by confounding variables. For the original (no confounding deflated) shapes (left), we found big differences due the inclu-
23 ouei h dutdmdli oepoone hni h unadjusted the in than pronounced more is model. model adjusted the ventricular prone the of of more in increase is base the volume adjusted it Finally, the so the thickening. stress, apex: myocardial mechanical of the bigger compensatory mass to in to exposed lower of is and and increase flatter base is The the LV in mass. concentrated is LV the model of in model increase difference unadjusted the large no while a found dilation, ventricular observed to we similar mass adjustment, myocardial re- confounder and different shape After base gave found inlet models the the sults. models however, than Both mass, more myocardial RV. dilated Regarding region and outflow apex. LV the ventricular both RV: the the for Both in of same changes adjust- increase the response. without an is measurement found and that models asociated with volumes unadjusted model their and PCA shows adjusted the 2.4 the by Figure obtained STD, ment. 2 + mean as did and redundant shape. became resulting adjustment the confounding influence the not (Figure of applied was use the deflation the of confounding minor the 2.2b), modes only When discriminating gave similar. most choice very resulting were DR the DR The dataset: variables. this confounding in the differences of not or sion after deflation. disappear confounding adjusment using to corrected subfig- due are In differences shapes the deflated. the that not were see are can method shapes we DR 2.2b, the the ure when to adjustment due model. differences the logistic that to the see secondary in can adjustment we confounding 2.2a, subfigure the In not or tech- using DR and of combinations niques different using obtained scores) between relation L 2.2: Figure 2 PCAPLS -Noconfounders rdc ewe h otdsrmntv hp onie ihtecor- the with coincides shape discriminative most the between product PCA -Noconfounders PLS -Noconfounders iue23dpcstems iciiaiesaepten,expressed patterns, shape discriminative most the depicts 2.3 Figure PCAPLS -Adjusted PCA -Adjusted PLS -Adjusted a rgnlshapes. Original (a)
L PCA - No confounders 2 PCAPLS - No confounders rdc ewe h otdsrmntv hp pattern(the shape discriminative most the between product L2 product PLS - No confounders
PCA - Adjusted
PCAPLS - Adjusted
PLS - Adjusted 0.75 0.80 0.85 0.90 0.95 1.00 PCAPLS -Noconfounders b ofudn deflation. Confounding (b) PCA -Noconfounders PLS -Noconfounders PCAPLS -Adjusted PCA -Adjusted PLS -Adjusted 24
PCA - No confounders
PCAPLS - No confounders L2 product PLS - No confounders
PCA - Adjusted
PCAPLS - Adjusted
PLS - Adjusted 0.75 0.80 0.85 0.90 0.95 1.00 Figure 2.3: The picture shows the mean shape of the population (right), and that mean shape after applying 2STD of the athletic remodelling ob- tained after adjusting by confounding (left) and the same remodelling with- out the adjustment (center). Both remodelling patterns show a dilation of the ventricles, with a bigger dilation of the RV outlet, but the one adjusted by confounders has a more pronounced dilation, and also a clear increase of the basal LV wall thickness. The different rows correspond to two dif- ferent views: the top row depicts a longitudinal view of the anterior wall of both ventricles and the bottom corresponds to a short axis view of the base, with the observer located in the atria.
Confounder adjustment No confounder adjustment
1.6 LV EDV LV Myocardial Mass RV EDV 1.4
1.2
1.0 Increase factor Increase factor
0.8
0.6 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 Remodeling score Remodeling score
Figure 2.4: The plot shows the measurement response to the remodelling: for each of the synthetic meshes generated by adding the shape remod- elling pattern to the mean shape with different magnitude, we compute classical measurements and show the variation ratio with respect to the mean shape.
25 Figure 2.5: Close-up on the right ventricular free wall of the different mod- els predicting the athletic remodelling shape pattern. The red-blue color map encodes the regional amount of remodelling: red means big dif- ferences compared to the control population, and blue small/ no remod- elling. We can see that PCA remodelling is smoothly distributed through the whole ventricle, while the results obtained with PLS and PCA + PLS present a more localized remodelling with a sharper red-blue transition: the remodelling is concentrated in the outflow.
The adjustment or not by confounders resulted in big differences be- tween the found remodelling patterns. On the other hand, differences due to DR were smaller, yet noteworthy. Figure 2.5 shows the remodelling pat- terns found for the different DR techniques, with a colormap showing the local amount of remodelling: reed indicates substantial changes and blue no remodelling. The figure shows a view of the RV free wall, which is the region that experienced more shape changes. All 3 shape patterns were similar and followed the same global trends, but they presented regional differences. We observed that both PLS and PCA + PLS found a remod- elling that was localised in the RV outlet and , at a lower degree, in the apex, while PCA showed a more spatially distributed remodelling that also affected the base and had a smoother transition between the affected and unaffected areas.
2.4.3. BMI model
To have a better understanding of the effect of an elevated BMI in the ventricles, we constructed a regression model that predicts BMI from the cardiac shape. As athletes and controls had different ranges of BMI, and to avoid finding any interference with the remodelling due to endurance sport, we only use the controls to build the BMI model. Also, since BMI
26 Figure 2.6: Synthetic representative shapes of a patient with BMI of 17.5 and another with BMI 30 according to the BMI predicting model. The re- modelling consisted of an increase of volume (specially in the axial direc- tions) and LV myocardial mass. and body size are related, we did not use the confounding variables in this model. The model was built using a PLS in its original regression mode, using 3 dimensions. To evaluate the model prediction capability, we com- puted the determination coefficient R2 using 5-fold CV, which was 0.44. Similar to classification, we obtained synthetic representative shapes (xˆb) associated with certain BMI value b. The synthetic mesh associated with a certain BMI was the one having minimal-distance to the mean shape (using the Mahalanobis distance), constrained to having the required pre- dicted BMI. The representative shapes associated to BMI values of 17.5 and 30, which are very extreme values, can be seen in Figure 2.6, and Figure 2.7 shows the measurement response of the remodelling. The BMI- associated remodelling consists of a moderate increase of ventricular size and a bigger increment of the myocardial mass.
27 LV EDV 1.3 LV Myocardial Mass RV EDV 1.2
1.1
1.0
0.9 Relative increment 0.8
0.7
17.5 20.0 22.5 25.0 27.5 30.0 32.5 BMI shape score
Figure 2.7: Measurement response of the BMI-related shape changes. As seen in the visual representation of the shapes, its main component is an increase of the myocardial mass, complemented with a smaller increase of volumes.
2.4.4. Confounding adjustment
Figure 2.8a depicts the L2 products between the athletic remodelling shape pattern (normalised to be a unitary vector) derived from the full population, and the shape pattern derived from the BMI-imbalanced pop- ulations. We found that without adjustment the imbalance confused the method, and it mixed the differences due to BMI with the ones related to endurance training. The adjustment by confounding variables resulted in a better agreement between the downsampled-derived remodelling patterns and the full-population pattern than the unadjusted model, as can be seen in Figure 2.9a. Figure 2.8b shows the dot product between the most dis- criminating shape pattern obtained with the downsampled (boxplot) and full (thick solid lines) datasets, and also the shape pattern associated to BMI, computed as described in section 2.4.3. There, we can see that in the downsampled datasets, the BMI-associated remodelling and the ath- letic one had lower correlation. This drop of correlation was larger for the models unadjusted by confounder variables. Figure 2.9a shows the measurement response to the athletic remod- elling of several models trained with a downsampled population. It shows different DR methods with confounding adjustment (upper row) and with- out confounding adjustment (lower row). We can see that when no con- founding correction is used, the found remodelling is associated with lower myocardial mass, contrary to what is observed with the full dataset ex-
28 Stability analysis Stability analysis 1.00 0.6
0.95 0.4
0.90 0.2
0.85 0.0
Dot product Dot product 0.2 0.80
0.4 0.75 Confounding adjustment Confounding adjustment No confounding adjustment 0.6 No confounding adjustment 0.70 PCA PCAPLS PLS PCA PCAPLS PLS
(a) Stability to imbalance (b) Dot product with BMI model.
Figure 2.8: Effect of using confounding adjustment in model stability from the imbalanced datasets for the 3 different DR methods. a) agreement of the model trained on downsampled data with the one on full data, mea- sured via its dot product. b) The thick solid lines correspond to the dot product between the BMI shape remodelling , and the athletic shape re- modelling derived from the full population. The boxplots show the dot product between the downsampled derived shape and the BMI shape re- modelling. When adjustment is used, the athletic and BMI mode have a positive relation (ie, they have remodelling partially in the same direction), due to both partially responding to an increase of pressure; however, that relation can become negative after downsampling, since controls become more overweight. periment in which myocardial mass was maintained. When the shape is adjusted by confounders, athletic remodelling was again associated with an increase of myocardial mass, consistent with the full population model. Figure 2.10 shows the discriminating shape patterns for a randomly se- lected downsampled dataset. There, we can observe that the model as- sociates athletic remodelling with a decrease of the wall thickness in the septum and apical regions of the LV when confounders are not considered. Resulting shape patterns were affected by the DR methods, PLS was more unstable than PCA and PCA + PLS, who had a better correlation to their full-population discriminating shape. In Figure 2.9a, we can see that PLS was not able to recover the increase of myocardial mass, and its mode presented no variation in the myocardial mass while both PCA and PCA + PLS could.
2.4.5. Confounding deflation
In this subsection we analyze the effect of the confounding deflation. As stated above, the shape prediction model of the confounding deflation
29 PCA + PLS PLS PCA
LV EDV 1.6 LV Myocardial Mass RV EDV 1.4
1.2
1.0 Increase factor Increase factor Increase factor
0.8
0.6 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 Remodeling score Remodeling score Remodeling score (a) Measurement response with confounder adjustment.
PCA + PLS PLS PCA 1.3 LV EDV LV Myocardial Mass 1.2 RV EDV
1.1
1.0 Increase factor Increase factor Increase factor
0.9
0.8
2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 Remodeling score Remodeling score Remodeling score (b) Measurement response without confounder adjustment.
Figure 2.9: Measurement response of the downsampled population with and without adjustment. We can see that the unadjusted methods find a negative relationship between athletic remodelling and LV mass, but the adjusted methods find a positive relationship. Figure 2.4 shows the equiv- alent plots for the models trained with the full population, which we use as groundtruth. The adjusted models are more similar to the groundtruth than the unadjusted. was trained using only the controls. We tested two possible scenarios: the case in which the athletes population was downsampled, and therefore the building of the shape residual model was not affected; and when the controls were downsampled (as in the previous subsection). The former was the most appropriate situation to apply confounding deflation, since there would not be extra unstability/bias introduced during the confound- ing deflation step, while the latter could introduce the bias in the dataset during the confounding deflation process. Finally, we show the potential danger of training the residual model using both populations and how the confounding deflation could even increment bias. When controls were not downsampled, the population used to train the shape prediction model was relatively unbiased, as is shown in Figure 2.11. Instead of downsampling the controls, we downsampled the ath- letes, removing athletes with high BMI analogously to the procedure used to downsample the controls. Figure 2.11a and 2.11b depict the same ex- periments as in the previous section: the dot product between the most discriminating shape obtained with the downsampled data and the remod- elling obtained considering the whole population, and also the dot product with the BMI mode. Results showed a considerable decrease in variabil-
30 Figure 2.10: Mean shape (right), and mean shape + 2STD of the athletic shape remodelling pattern(left), derived from a BMI-imbalanced popula- tion. The pattern in the center column was obtained without adjustment, while the one on the right was adjusted. We can that the unadjusted model finds a decrease of myocardial mass in the apical and septal walls, while the adjusted finds an increase of mass ity and a higher correlation with the full-dataset-derived remodelling than the confounder adjustment experiments (Figure 2.8a). Adding confounder adjustment on top of confounding deflation did not produce an increase of accuracy. Figure 2.13a shows the results when the population in which the resid- ual is trained is downsampled. We downsampled the controls based on their BMI. Results were much worse than when athletes were downsam- pled, and even worse than a simple confounder adjustment. Adding con- founding adjustment on top of confounding deflation had a beneficial ef- fect. Figure 2.13b shows the effect of using both athletes and controls in the training of the shape-prediction model for confounding deflation: there was a drop in stability compared to the use of a single population when PLS and PCA + PLS were used (Figure 2.8a). Strangely there was an im- provement compared to the baseline (where confounders are completely ignored). Therefore we observed that using both populations in the training of the shape prediction model during the confounding deflation step resulted in worse results. We explored the reason why using both populations can
31 Stability analysis Stability analysis 1.00 0.8
0.95 0.6
0.90 0.4
0.85
0.2 Dot product Dot product 0.80
0.0 0.75 Confounding adjustment Confounding adjustment No confounding adjustment No confounding adjustment 0.70 PCA PCAPLS PLS PCA PCAPLS PLS
(a) Dot product with full data (b) Dot product to BMI model for model. residual.
Figure 2.11: Stability analysis of the confounding deflation method when the athletes (that are not used in the construction of the residual model) are downsampled.
Full population Downsampled + Conf Downsampled + No Conf 1.6 LV EDV LV Myocardial Mass RV EDV 1.4
1.2
1.0 Increase factor Increase factor Increase factor
0.8
0.6 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 2.0 1.5 1.0 0.5 0.0 0.5 1.0 1.5 2.0 Remodeling score Remodeling score Remodeling score
Figure 2.12: Measurements response of the shape pattern found with the athlete-downsampled data. There were very small differences between the different DR methods, and all methods were able to recover the original shape pattern. create a confounding effect, associating an imbalance in a variable to the inter-class shape differences. This can happen even when the variable is not associated to any shape remodelling. To illustrate this effect, we cre- ated a dummy synthetic variable which is just the athlete label plus Gaus- sian noise. Obviously, by generation we knew that this variable did not have any direct relationship to shape. To evaluate this effect, we computed the L2 dot product with the most discriminating shape between athletes and controls and the shape pattern associated to the dummy variable in the shape-prediction model. We repeated this full variable generation and L2 products computation process 100 times to remove randomness of the analysis. Figure 2.14 (left) shows the distribution of this dummy variable for a certain seed. We constructed two residual models one only with the controls, and the other with both populations. Figure 2.14 (right) shows the distribution of the dot product of the shape prediction model coeffi- cients associated to the dummy variable when both populations are used, and when only the controls are used. The shape associated to the dummy
32 variable is independent to the athletic remodelling when only one popu- lation is used, but becomes very similar to the control-athlete difference when both populations are used in training due to confounding effect.
Stability analysis Stability analysis 1.00 1.00
0.95 0.95
0.90 0.90
0.85 0.85 Dot product Dot product 0.80 0.80
0.75 0.75 Confounding adjustment Confounding adjustment No confounding adjustment No confounding adjustment 0.70 0.70 PCA PCAPLS PLS PCA PCAPLS PLS
(a) Confounder deflation trained (b) Confounder deflation trained on controls only. on both classes.
Figure 2.13: Effect of the population used to train the confounder deflation model on the discriminative pattern stability, assessed via its L2 product with the result obtained using the full population. Subfigure a) shows when the training of the shape prediction model in the confounder deflation step is trained using the downsampled class, and subfigure b) shows when both controls and athletes are used for training.
Figure 2.14: This figure shows the distribution of the dummy variable used as a confounder, and the dot product between the regression coefficients associated to the dummy variable in the residual model and the athletic remodelling shape pattern.
2.5. Discussion
The shape models corresponding to the athletic and overweight re- modelling corroborated the current clinical literature. The BMI remodelling consisted in mostly a concentric remodelling, both the LV mass and LV EDV increased, increasing more the mass than the volume: this coincides with the hypothesis that remodelling is concentric to cope with elevated
33 pressure (increase of myocardial mass) and CO demands (increase of volume). Another hint that the remodelling was mostly pressure driven is that the septum of the mesh representing a high BMI individual presents a bulge below the aorta, that has been described as an early indicator of elevated pressure (Baltaeva et al., 2007; Gaudron et al., 2016). The athletic remodelling consisted of a predominant increase of volume, and if confounder adjustment was performed, an increase of myocardial mass. There is some controversy if endurance athletes remodelling is more ec- centric ( they increase their volume more than their mass) or concentric (they increase their mass more than their volume), but there is a consen- sus that there is both an increase of mass and volume (Scharhag et al., 2002), and therefore we considered the adjusted model more accurate. The increase of ventricular volume in the confounder adjusted model is considerably larger than the unadjusted. These differences between the adjusted and not adjusted models can be explained because the variability in size and myocardial mass can come also from the patients morphomet- rics: big persons have big hearts. Without considering this extra non- imaging information, it is not possible to discern when the size of the heart is due to the patient being big, or the heart had a dilation as a remodelling reaction to exercise. The artificially generated imbalanced sets allowed us to validate our hypothesis: without any correction, the athletic remodelling presented a reduction in LV myocardial mass. This is obviously false, since all stud- ies have found that endurance exercise provokes an increase of myocar- dial mass. The downsampled control population had a high percentage of overweight, who had a concentric remodelling and causing the previ- ous association with control and higher myocardial mass was done. The stability analysis, also showed a bigger match between the downsampled- derived remodelling and the full-dataset one when confounding adjustment was used. We also studied another strategy: confounding deflation. Confound- ing deflation consists of generating a model that predicts shape from the confounders, and working with the residual of that prediction. Our results showed that this strategy worked well when there was access to a good population to train the shape prediction model: otherwise it can actually increase bias. This was shown when training on the downsampled popu- lation, or when both the case and control were used simultaneously. Finally, we compared different linear methods for DR: PCA and PLS. In the original dataset, we obtained a better classification accuracy using PLS. We also showed that PLS was able to capture more localised re- modelling than PCA, that was limited to smooth and global remodelling.
34 However, PCA overperformed PLS in stability tested during the downsam- pling analysis. Both models could be combined by using a coarse PCA, that only removes the modes encoding very little variance, with a PLS. With this method we were able to capture both localised remodelling and have high stability to imbalances.
2.6. Conclusion
We have presented a SSA framework to find regional shape differences between two populations, taking special care to correct for any potential bias related to demographics parameters. The framework is fully linear: it used a PCA sequentially combined with a PLS as dimensionality reduc- tion of the shapes, followed by a logistic regression model. The linearity allows to easily interpret and visualise the model and build synthetic rep- resentative shapes of the model. To correct for confounding effects, it in- corporates adjustment, where the confounding variables are added to the logistic model and confounding deflation, which consists of building a re- gression model that predicts shape from confounding variables, and we used it to remove the shape variability associated to these variables. We applied our framework to a real dataset consisting of athletes and controls to find the remodelling due to the practice of endurance exer- cise. Our results confirmed the current literature on endurance-sport re- modelling in the LV: ventricular dilation and increment of myocardial mass, specially in the basal area. In the RV, we found that the volume increase was not homogeneous but concentrated in the outflow. In the controls, we used an adaptation of our classification framework to regression to explore obesity remodelling and found it to be mainly an increase of myocardial mass. In this population, we analysed the effect of confounders in a semi- synthetic dataset obtained by downsampling the control population non- uniformly, keeping individuals with high BMI. Even if athletic remodelling is very prominent, we were able to bias the model to output that athletes have lower myocardial mass than controls. This was corrected when adjustment was used. However, we found that we could only use confounding deflation when the control population was relatively big and balanced, and if that is not the case using confounding deflation can actually increment bias. In our work we have tested only linear SSA methods, but this confounder- related problems might appear even more with the use of more complex frameworks, able to capture non-linear shape patterns to capture subtler morphology differences in populations that are almost indistinguishable from controls and differences cannot be found through traditional means.
35
Chapter 3
VOLUMETRIC PARCELLATION OF THE RIGHT VENTRICLE FOR REGIONAL GEOMETRIC AND FUNCTIONAL ASSESSMENT
3.1. Introduction
When altered stimuli (such as pressure or volume loading) are present, cardiac morphological remodelling is induced through the ability of the in- dividual cardiac myocytes to change size and shape (Arts et al., 1994; Grossman et al., 1975; Opie et al., 2006). This remodelling can either be adaptive, in which it helps the heart to compensate for the changes and keep pumping enough blood to satisfy the system oxygen demands while at the same time maintaining pressure within physiological range. Or, it can be maladaptive in which case the changes are not compensated thus provoking a maladjustment that starts damaging the heart or makes it un- able to satisfy the systemic oxygen demands and fail. In clinical practice, remodelling of a ventricle is often simplified to global changes. These are categorised as either wall thickening with inward mo- tion of the inner wall (thus reducing cavity size and wall stress) as a re- action to pressure loading - often referred to as concentric remodelling, or This chapter is adapted from: Bernardino G, Hodzic A, Langet H, González Ballester M.A., De Craene M., Saloux S.,Bijnens B.. "Volumetric parcellation of the right ventricle for regional geometric and functional assessment." [In preparation]
37 dilatation of the cavity to cope with volume overload so that, without chang- ing the wall deformation during contraction, more stroke volume is ejected with each beat - often referred to as eccentric remodelling. An example of a maladaptive remodelling would be the hypertrophic cardiomyopaty (HCM) (Olivotto et al., 2012), where the concentric remodelling thickens the myocardial walls so much that the filling volumes are reduced, lead- ing to diastolic heart failure, and the outflow tract can become obstructed. Maladaptive remodelling can get the heart into a vicious circle in which it tries to maintain blood flow by contracting harder, increasing even more the pressure, thus triggering more concentric remodelling. Therefore, it is clinically very important to assess the induced changes and interpret them to distinguish different types of remodelling patterns. The underlying mechanism of remodelling is the adaption of individual myocytes as response to very local stimuli. Additionally, the cardiac cham- bers are no spheres and the muscle is highly anisotropic with internally changing myocardial fiber orientations depending on the position within the wall. Therefor, changes induced by alterations in loading conditions do not have the same effect in all regions of the heart but show specific local trends, such as for instance the presence of a basal septal bulge in- duced by hypertension (Baltaeva et al., 2007) or a base-to-apex gradient in deformation in deposition diseases as amyloidosis (Cikes et al., 2010) or thalassemia (HAMDY, 2007). However, in clinical practice, these regional patterns are often ignored and the quantitative assessment of morpholog- ical remodelling is performed by interpreting only global measurements, namely ventricular volumes and myocardial mass (particularly in the case of the left ventricle (LV)). There have been some efforts and advances towards the quantifica- tion of regional patterns although these mostly involve segmental motion assessment (particularly in coronary artery disease) rather than local ge- ometry. For the LV, a standardised partition in 17 wall segments has been proposed (Cerqueira et al., 2002), which recently also has been used to quantitatively assess regional strain patterns. However, these wall seg- ments are by definition equal in size and thus of limited utility to assess changes in morphology. Compared to the LV, the right ventricle (RV) has a complex and ir- regular shape, with more variability (Haddad, Doyle, et al., 2008) so that regional analysis of the RV is even more difficult than of the LV. While there have been studies referring to a regional analysis of the RV (Adde- tia, Maffessanti, Yamat, et al., 2016; Addetia, Maffessanti, Muraru, et al., 2018; Moceri et al., 2018), these are limited to the epicardium, and mostly focus on assessing wall curvature. While local curvature is an important
38 component of the wall stress generated by pressure and thus very im- portant in pressure-overload remodelling, an analysis in terms of regional dilatation/volume is needed to correctly assess remodelling from volume- overload. Therefor, there is a need for quantitative approaches that can assess regional shape as well as volume remodelling in a clinically relevant and physiological plausible way. To address this, in medical image analysis, the typical method to as- sess regional morphology is through the creation of an atlas, which is a template shape representative of a population, and registering each patient-specific shape to this atlas. This has shown useful to describe inter-individual variations overall morphology in populations, but is more challenging when following subtle regional remodelling within an individual over time. Additionally, this approach has the drawback that it requires reg- istration (deforming the atlas to match the individual) (Joshi et al., 2004). This is an unstable and computationally expensive process. Especially since there are few and separated landmarks for the ventricles, this regis- tration is based on image intensity/shape patterns, and has no guarantee of the correctness of the point-to-point correspondence. Thus, when using computational meshes to represent the heart through image segmenta- tion, after atlas registration, an important part of the mesh nodes’ positions do not correspond to identifiable anatomic landmarks among different in- dividuals so that point-to-point correspondence cannot be used anymore to assess physiological remodelling. To avoid explicit registration, some authors have proposed parametri- sation methods to create anatomical maps of organs(Nuñez-Garcia et al., 2019; Vera et al., 2013; Paun et al., 2017) by finding smooth bijective maps 2 3 from each surface/volume to a common domain, a subset of R or R re- spectively. The parametrisations of individual anatomies can subsequently be used to obtain a point-to-point correspondence. The mapping is typi- cally obtained through a minimisation of some kind of distortion metric. The most common is trying to force it to be as conformal as possible, ie that locally maintains angles. This is typically enforced by the minimisation of a Laplacian energy (resp. Laplace-Beltrami), but it is not the only pos- sibility: there are other approaches that try to maintain, for instance, local distances (Sorkine et al., 2007). In this chapter we develop this further for a clinical application, using images of suboptimal quality, to propose an automatic method for mesh- independent volumetric parcellation of the RV based on the geodesic dis- tance to three easily identifiable landmarks: tricuspid valve, pulmonary valve and apex. To assess the method, to identify clinically relevant and
39 physiologically plausible regional remodelling, we validated is using a syn- thetic dataset created through regional induction of circumferential and lon- gitudinal elongation and analyse the sensitivity of the parcellation to both global and regional. To assess the performance, and robustness to noise, in a real setting, we also do an inter- and intra- observer reproducibility analysis, as well as a test/retest comparison of two sequential acquisitions, on the same patient. This technique enables both a regional analysis of anatomy, using the end diastolic (ED) volumes, as well as function, via regional ejection fraction (EF).
3.2. Methodology
3.2.1. Data acquisition
3D echocardiographic images of 5 competitive american style foot- ball players were acquired using a modified apical 4 chambers view us- ing an EPIQ7 ultrasound system (Philips Medical Systems, Andover, MA) equipped with a 1 to 5 MHz transthoracic matrix array transducer (X5-1). For each patient, 4-6 different ECG-gated subvolumes where acquired in a single breath-hold to be compounded into the full 3D+t images of 2 com- plete cardiac cycles. A written informed consent form was obtained from all study participants. A control was imaged in two consecutive acquisitions by different op- erators to obtain an estimate on the variability due to the imaging process. The image loops were processed using a clinically validated software (4D RV-FUNCTION Tomtec-Arena TTA2, Tomtec Imaging Systems GmbH, Unterschleissheim, Germany (Niemann et al., 2007)) to segment the RV and obtain a 3D-model for each patient. These models were exportable in ucd, a standard file format. All 3D models had the same topology: a tri- angular watertight mesh with 938 nodes and 1872 faces. The points were in approximate point-to-point correspondence. The segmentation pipeline consisted of: the clinician segments the RV endocardium contour of the frame corresponding to the R peak using the semiautomatic tool, and the result is tracked during a full cycle. Afterwards, the clinician can adjust the end systolic (ES) and ED (defined as the peak of the R wave in the ECG) segmentation iteratively until visually satisfied with the resulting contours.
3.2.2. Parcellation of the right ventricle
Compared to the LV, which resembles a prolate ellipsoid, the RV has a more complex shape and partially surrounds the LV. Its anatomy is most
40 commonly described biaxial: one axis goes from tricuspid valve to the apex, and the other from the apex to the pulmonary valve. The RV can be grossly separated in 3 main anatomically and functionally different parts: the outflow infundibulum, the smooth inlet and the trabeculated apex (Had- dad, Doyle, et al., 2008; Haddad, Hunt, et al., 2008). However, there is no consensus on the exact border between these parts, and different experts can draw different partition boundaries over the sames images. Given that the partition is crucial to described volume and shape chan- ges over time when doing follow-up in individuals, we propose an auto- matic method for volumetric parcellation of the RV, based only on geomet- ric properties. This partition has the advantage of being fully automatic and therefore completely reproducible under the same image and seg- mentation, however it still depends on image and segmentation quality. To avoid errors due to a bad point-to-point correspondence, our parcella- tion only uses the geodesic distances from anatomic landmarks that can be clearly identified in 3D echocardiography: the apex, tricuspid and pul- monary valve. The method is independent of the exact triangulation of the ventricular surface. We applied our method to the analysis of 3D echocardiography images of the RV processed with Tomtec software, but it can easily be adapted to other processing platforms and imaging modalities. Figure 3.1 shows the full process used to parcellate the RV. The first step of our parcellation is the identification of the valves and apex using the provided point-to-point correspondence in the surface mesh (given that these are stable anatomical landmarks in the mesh). Next, for each node of the mesh, we compute the geodesic distances to the apex, pulmonary valve and tricuspid valve. The geodesic distances between two points are computed on the surface with an exact algorithm (Surazhsky et al., 2005) that computes the length of the minimal on-surface path between two points. The distance between a point and an anatomical substructure is defined as the minimum distance from the point to any point that belongs that anatomical substructure:
dt(x) = min{dgeo(x, y)|y ∈ tricuspid valve} dp(x) = min{dgeo(x, y)|y ∈ pulmonary valve} (3.1) da(x) = min{dgeo(x, y)|y ∈ apex}
Figure 3.1.a shows the geodesic path from a sample point to the dif- ferent landmarks, and Figure 3.1.b shows distance from every point of the surface to the apex represented as a heatmap: the points furthest to the apex are coloured in red and the closest in blue. After this distance is com- puted for every point of the surface, the interior of the triangular surface
41 (a) (b) (c)
(d) (e) (f)
Figure 3.1: Steps to generate the volumetric partition. a) For each point, the geodesic distances to the apex/tricuspid and pulmonary valve are com- puted. b) The geodesic distances to each of the landmarks define a scalar map over the surface mesh. c) This distance map from the surface to the cavity by tetrahedralising the mesh, and the Laplace equation is used to interpolate values to the interior. d) The ventricle is split in the regions by assigning each point of the cavity to the closest landmark. e-f) Visualisa- tion of the RV parcellation over slices of the original 3D images.
42 mesh is tetrahedralissed using a publicly available software (TetGen ver- sion 1.5.1, (Si, 2015)) (Figure 3.1.c). The distance defined on the surface of the mesh is propagated to the interior of the ventricle using Laplace’s equation. This equation is discretized using finite elements with a publicly available software (Sfepy (Cimrman et al., 2019)). The equation uses the tetrahedralized mesh as domain and Dirichlet boundary conditions speci- fied by the surface-defined distance maps. Formally, this interpolation step is defined as follows, where M ∈ {apical, inlet, outflow}, Ω refers to the volumetric domain and ∂Ω to its boundary (the surface mesh):
∆u = 0 for x ∈ ˚Ω M (3.2) uM (x) = dM (x) for x ∈ ∂Ω
This process is repeated to compute and extend to the interior of the cavity of the 3 distances. Once the distances are defined in the volumet- ric mesh, we partition the ventricle, assigning each point of the mesh to the "closest" landmark, using the interpolated distances. Minlet,Mapical and Moutlet respectively are the partition corresponding to the inlet, apex and outflow. Each point is assigned to the region whose representing land- mark is "closer", using the interpolated dzs, as shown in Figure 3.1.d. The partition does not follow the mesh vertices and edges, but new elements are generated during the partition. We used linear interpolation to define the distance values inside each tetrahedron. A formal definition of the seg- ments is:
Minlet = {x|dtricuspid (x) < dpulmonary (x) , dtricuspid (x) < dapex (x)}(3.3)
Moutlet = {x|dpulmonary (x) < dtricuspid (x) , dpulmonary (x) < dapex (x) (3.4)
Mapical = {x|dapex (x) < dtricuspid (x) , dapex (x) < dpulmonary (x)(3.5)
This partition can be propagated from the ED-surface to ES-surface using the point-to-point correspondence between surfaces belonging to the same individual, that are obtained via tracking the initial surface, and then extended to the interior cavity via the same Laplacian interpolation. With this procedure, we can compute regional ES volumes and ejection fractions, allowing for regional functional assessment of the acRV.
3.2.3. Local and global anatomic frame of reference
To clinically interpret local geometric changes, it is better to work in an anatomical frame of reference, with longitudinal and circumferential direc- tions, instead of the Cartesian system of coordinates. At each point of the
43 Figure 3.2: Circumferential (right) and longitudinal (left) directions defined in each triangle of a sample RV mesh. The tricuspid valve is shown in green and the pulmonary in blue. Since valve are not part of the my- ocardium, the definition of the anatomical direction there has no meaning. mesh, circumferential and longitudinal directions are defined locally using the method proposed by Doste et al., 2019. We defined the longitudinal direction using the stationary heat flow in surfaces, with a cold source in the apex, and two hot sources at the same temperature in the two valves. The heat flow is solved using the Laplace-Beltrami linear differential equa- tion on a surface. The Laplace-Beltrami operator (∆) is discretized using the cotangent formulation (Pinkall et al., 1993):
∆u = 0 u(apex) = 0 (3.6) u(valves) = 0
The longitudinal direction (l) at each point is the result of normalising the resulting temperature gradient. The circumferential (c) is chosen to be orthogonal to both the longitudinal and the surface normal at that point (n), so that (l,c) form a base of the tangent space at the given point:
∇u l = (3.7) kuk c = l × n (3.8)
Figure 3.2 shows the local circumferential and longitudinal directions. A global longitudinal direction is computed by averaging all the local ones, and the circumferential directions are defined as orthogonal to the longitu- dinal.
44 3.2.4. Strain as a value to express local deformation
For two meshes, a reference and a deformed mesh, in point-to-point correspondence, we can compute the strain associated to the deforma- tion. This strain fully characterises the deformation, modulo rigid body deformations. To interpret this strain, it is more natural to work in the previ- ously defined local anatomical reference frame. For each triangle, we can express each of its edges as a combination of the local anatomical direc- m tions (l, c). We note Et the concatenated vectors corresponding to the edges of triangle t and mesh m. With this, we can compute the local linear transformation Ft at triangle t. In a continuous setting, this Ft corresponds to the Jacobian matrix of the deformation.
mdef mref −1 Ft = Et Et (3.9)
From this transformation, we can compute the Cauchy strain tensor for small displacements = (F t + F )/2, and extract the strain in the longi- tudinal (ll) and circumferential (cc) directions. Note that longitudinal and circumferential directions are not necessarily aligned to the principal strain directions, which are the eigenvectors of the strain tensor.
3.2.5. Synthetic regional remodeling patterns generation
For the localised remodelling, we used a modification of a linear surface construction method (Wang et al., 2012) that generates 3D meshes from a local description. The generated mesh has a fixed topology, that cannot contain a non-manifold edge. The local description consists of the edge lengths and the dihedral angle associated to every edge (the angle formed by the normals of the adjacent triangles). Obviously, not all combinations of lengths and angles define a valid surface, but we can formulate the reconstruction in a minimisation setting so we obtain the possible surface satisfying as much as possible the local description.
Variables definition
We describe a local frame of reference for each triangle of the mesh from the input data. This frame is arbitrary but uniquely defined, assuming a unique ordering of the nodes inside each triangle. It uses the first node of the triangle as origin, the direction of the first edge as x-axis, the normal as the z-axis and then completes the base to be orthonormal. We will t call ai the coordinates of the i-th point of the triangle expressed in local t frame. By convention, a0 = (0, 0, 0), the third coordinate is always 0 for
45 Figure 3.3: Two adjacent triangles Ti and Tj, with their respective local systems of references. dij is the dihedral angle, which is the angle formed between the two normals. It is represented in the midpoint of the common edge. fi and fj are the orthogonal frame of reference associated to each triangle.
t all points (since they are coplanar), and a1 = (x, 0, 0) . Note that by basic trigonometry we can compute the local coordinates in that of the triangle nodes given the 3 lengths, using the constraints that the first node is in 0, and the second one lies in the x-axis. The method unknown variables are the 3D coordinates xi for each mesh point i, and a reference frame ft associated to each triangle t, that corresponds to the mapping from the triangle coordinates a to the 3D space. Figure 3.3 shows two adjacent triangles, with their dihedral angle and the associated frames of reference. For two adjacent triangles i and j, we can obtain the rotation Rij from frame fj to fi: T Rij = fj fi (3.10)
We can express Rij using only elements of the local descriptor: the θ triangle coordinates and the dihedral angles dij. We call φv to the rotation of angle θ around the axis of rotation v. Let e be the common edge between triangles Ti and Tj, we can compute the angle θ between edge e and the 0 first edge of Ti, and θ is the respective angle for Tj. Then, we can express Rij as the composition of 3 rotations:
θ dij θ Rij = φzφx φz (3.11)
Linear reconstruction
The reconstruction method is an inverse problem of the computation of the local triangle coordinates as and transition matrices Rs. We operate
46 the previous equations to obtain an equivalent form linear on xs and fs, and use their quadratic residual as energy, thus formulating an optimisa- tion problem with the world position of the nodes x and the frames f as variables. To easily solve this problem, we do not enforces that matrices fs are rotations, but general matrices. By multiplying by fj equation 3.10 and rearranging terms, we obtain the following equivalent equation:
fi − fjRij = 0 (3.12)
For every edge eij in every triangle t, where the nodes have in-triangle indices i’ and j’ respectively, we can use that ft to transform from triangle coordinates to world coordinate: xi − xj = ft ati0 − atj0 (3.13)
After moving all terms to the LHS, we obtain: