Lung Lobe Segmentation Based on Lung Fissure Surface Classification

algorithms

Article Lung Lobe Segmentation Based on Lung Fissure Surface Classiﬁcation Using a Point Cloud Region Growing Approach

Xin Chen 1, Hong Zhao 2 and Ping Zhou 3,*

1 College of Electrical and Information Engineering, Hunan University, Changsha 410082, China; [email protected] 2 College of Aerospace Science and Engineering, National University of Defense Technology, Changsha 410003, China; [email protected] 3 College of Biology, Hunan University, Changsha 410082, China * Correspondence: [email protected]

Received: 2 September 2020; Accepted: 13 October 2020; Published: 15 October 2020

Abstract: In anatomy, the lung can be divided by lung fissures into several pulmonary lobe units with specific functions. Identifying the lung lobes and the distribution of various diseases among different lung lobes from CT images is important for disease diagnosis and tracking after recovery. In order to solve the problems of low tubular structure segmentation accuracy and long algorithm time in segmenting lung lobes based on lung anatomical structure information, we propose a segmentation algorithm based on lung fissure surface classification using a point cloud region growing approach. We cluster the pulmonary fissures, transformed into point cloud data, according to the differences in the pulmonary fissure surface normal vector and curvature estimated by principal component analysis. Then, a multistage spline surface fitting method is used to fill and expand the lung fissure surface to realize the lung lobe segmentation. The proposed approach was qualitatively and quantitatively evaluated on a public dataset from Lobe and Lung Analysis 2011 (LOLA11), and obtained an overall score of 0.84. Although our approach achieved a slightly lower overall score compared to the deep learning based methods (LobeNet_V2 and V-net), the inter-lobe boundaries from our approach were more accurate for the CT images with visible lung fissures.

Keywords: lung lobe segmentation; lung fissure classification; point cloud; region growing; normal vector and curvature; principal component analysis; multistage spline fitting

1. Introduction The human lung is divided into five lobes by the lung fissures and lung boundaries. The lung lobe is a relatively independent unit with regards to anatomical structure and function. Knowledge of the distribution of lung diseases among the lung lobes is important for treatment planning [1]. Moreover, lung lobe segmentation is indispensable for quantitative evaluation of lung diseases. Manual lung lobe segmentation suffers from subjective bias and variance. Therefore, there is a need to develop an automatic lung lobe segmentation method with high robustness and accuracy for clinical image data analysis. Presently, a number of works have been devoted to lung lobe segmentation, which can be divided into two categories: the supervised machine learning method and the unsupervised traditional medical image processing method according to [2]. The traditional segmentation methods can be further classified into two categories: the algorithm based on the fusion of anatomical structure information and the method without the aid of anatomical structure.

Algorithms 2020, 13, 263; doi:10.3390/a13100263 www.mdpi.com/journal/algorithms Algorithms 2020, 13, 263 2 of 12

Some machine learning methods have been used to segment lung lobes in recent years. George et al. usedAlgorithms progressive 2020, 13 holistically, x FOR PEER REVIEW nested neural networks to predict lung lobe boundaries in22D of 12 slices, and then combinedSome machine them learning with the methods 3D random have been walker used to to achieve segment lung lung lobe lobes segmentation in recent years. [3]. George Imran et al. proposedet al. used a fast progressive and reliable holistically progressive nested denseneural V-networknetworks to [predict4]. In lung order lobe to boundaries prevent over in 2D fitting, Ferreiraslices, et and al. then integrated combined several them with regularization the 3D random methods walker to on achieve the basis lung of lobe a residualsegmentation network [3]. [5]. TheseImran methods et al. proposed based on a deepfast and learning reliable can progressiv improvee dense the eV-networkfficiency of[4]. lung In order lobe to segmentation prevent over to a certainfitting, extent, Ferreira but et they al. integrated need a large several amount regularization of training methods data toon bethemanually basis of a residual labeled network by experts [5]. and high-endThese Graphics methods Processingbased on deep Unit learning (GPU) can equipment improve th fore training.efficiency of lung lobe segmentation to a certainIn the extent, unsupervised but they traditionalneed a large lungamount lobe of trai segmentationning data to algorithms,be manually labeled there are by experts two main and steps includinghigh-end lung Graphics segmentation Processing and Unit lung (GPU) fissure equipment detection. for training. The former has been fully studied and is a In the unsupervised traditional lung lobe segmentation algorithms, there are two main steps stable and reliable algorithm [6–8]. On the other hand, automatic detection of pulmonary fissures is including lung segmentation and lung fissure detection. The former has been fully studied and is a still challenging.stable and reliable There algorithm are some [6–8]. methods On the that other use hand, the localautomatic and globaldetection knowledge of pulmonary of lung fissures anatomical is structuresstill challenging. to locate the There lung are fissure some methods area, such that as use using the local airway and andglobal vessel knowledge information of lung [anatomical9–12], which is basedstructures on the assumption to locate the lung that therefissure are area, no such blood as using vessels airway penetrating and vessel the information pulmonary [9–12], fissure which and the airwayis based tree canon the be assumption divided into that branchthere are trees no blood belonging vessels penetrating to five lung the lobes. pulmonary These fissure methods and the require segmentationairway tree of can the be airway divided and into vessel branch tree, trees so thebelonging algorithm to five is time-consuming,lung lobes. These methods and its performancerequire dependssegmentation on segmentation of the airway accuracy. and ve Forssel pathological tree, so the algorithm data, once is time-consuming, a segmentation errorand its occurs, performance detection of the pulmonarydepends on fissure segmentation cannot beaccuracy. carried For out patholog successfully.ical data, once a segmentation error occurs, detection of the pulmonary fissure cannot be carried out successfully. Because the right lung has two types of fissures, the horizontal fissure and the oblique fissure, it is Because the right lung has two types of fissures, the horizontal fissure and the oblique fissure, it necessary to classify them before segmenting the lobes correctly. In this paper, a lung lobe segmentation is necessary to classify them before segmenting the lobes correctly. In this paper, a lung lobe methodsegmentation based on lungmethod fissure based classification on lung fissure is proposed. classification In orderis prop toosed. optimize In order the calculationto optimize andthe save space,calculation we represent and save the pulmonary space, we represent fissure surface the pulm extractedonary fissure by a directionalsurface extracted derivative by a ofdirectional plate (DDoP) filterderivative [13] as the of pointplate (DDoP) cloud data filter structure.[13] as the point Based cloud on this, data westructure. can calculate Based on the this, normal we can and calculate curvature differencesthe normal of the and pulmonary curvature differences fissure surface of the with pulmon highary accuracy, fissure surface and then with realize high accuracy, the pulmonary and then fissure classificationrealize the using pulmonary a region fissure growing classification approach. using a For region missing growing pulmonary approach.fissures For missing after pulmonary classification, theyfissures are interpolated after classification, and expanded they byare theinterpolated multistage and spline expanded surface by fitting the multistage method, andspline then surface combined fittedfitting fissures method, with and a lung then combined mask to realizefitted fissures lung with lobe a segmentation. lung mask to realize This lung method lobe segmentation. does not rely on This method does not rely on anatomical structure information such as airways or vessels, and anatomical structure information such as airways or vessels, and experiments show that it has higher experiments show that it has higher reliability for detecting mild or moderate lesions on lung CT reliability for detecting mild or moderate lesions on lung CT images. images. 2. Lung Lobe Segmentation Algorithm 2. Lung Lobe Segmentation Algorithm FigureFigure1 shows 1 shows the the basic basic anatomical anatomical structures structures ofof the human lung. lung. Based Based on onthis this anatomical anatomical information,information, we we propose propose aa three—step lung lung lobe lobe segmentation segmentation framework, framework, which includes which (1) includes lung (1) lungsegmentation, segmentation, (2) (2) lung lung fissure fissure classification, classification, and and(3) lung (3) lung fissure fissure surface surface fitting fitting and lung and lunglobe lobe segmentation.segmentation. Among Among them, them, we we adopt adopt aa relativelyrelatively mature mature method method for for lung lung segmentation, segmentation, which which is is brieflybriefly summarized. summarized. The The other other steps steps are are moremore thoroughly described described in inthis this section. section.

FigureFigure 1. 1. CTCT image of of human human lung. lung. Algorithms 2020, 13, x FOR PEER REVIEW 3 of 12

Algorithms 2020, 13, 263 3 of 12 2.1. Lung Segmentation In order to segment lung lobes, it is necessary to correctly segment and identify the lungs including2.1. Lung Segmentationthe left lung and right lung. We adopted the method proposed in [14], which mainly includedIn order coarse to segment airway lungsegmentation, lobes, it is necessarylung segmen to correctlytation, segmentseparation and of identify left and the right lungs lung including and pulmonarythe left lung andholes right filling lung. four We adoptedsteps. Firstly, the method the proposedairway was in [14 located], which mainlyautomatically included by coarse the morphologicalairway segmentation, method lung after segmentation, threshold processing separation in ofseveral left and cross-sectional right lung and slices pulmonary at the top holes of the filling CT imagefour steps. and segmented Firstly, the airwayby using was the located region automaticallygrowing method by thewith morphological leakage control, method where after voxels threshold at the airwayprocessing were in taken several as cross-sectionalseed points. Secondly, slices at thethe topregion of the growing CT image method and segmented was further by used using to the segment region thegrowing lung. methodThen, after with the leakage removal control, of the where airway, voxels the atseparation the airway of werethe left taken and as right seed lung points. was Secondly, judged bythe connectivity region growing analysis. method If not was successful, further used the corrosion to segment operator the lung. was Then, run iteratively after the removal until the of two the wereairway, separated. the separation Finally, of by the morphological left and right lung closing was operation, judged by we connectivity filled the vessel analysis. holes If not to obtain successful, the leftthe and corrosion right lung operator masks. was run iteratively until the two were separated. Finally, by morphological closing operation, we filled the vessel holes to obtain the left and right lung masks. 2.2. Pulmonary Fissure Classification 2.2. Pulmonary Fissure Classification The fissure formed by the bilayer visceral membrane in the lung is called the pulmonary fissure, whichThe divides fissure the formed lung into by the several bilayer lobes, visceral as shown membrane in Figure in the 1. lung After is calledobtaining the pulmonarythe left and fissure, right lungs,which further divides pulmonary the lung into fissure several extraction lobes, as shownand classification in Figure1. are After needed. obtaining Firstly, the leftwe andused right the DDoP lungs, filterfurther and pulmonary directional fissure partition extraction denoising and classification and fusion arealgorithm needed. to Firstly, enhance we used and thesegment DDoP the filter lung and fissuredirectional region partition in the denoisingCT image. and Then, fusion we algorithmused the topoint enhance cloud and to represent segment the the lung sparse fissure lung region fissure in surface,the CT image. and finally Then, we we clustered used the the point lung cloud fissure to represent according the to sparse the differences lung fissure in surface,the normal and vector finally andwe clusteredcurvature the on lungthe lung fissure fissure according surface. to the differences in the normal vector and curvature on the lung fissure surface. 2.2.1. Pulmonary Fissure Segmentation 2.2.1. Pulmonary Fissure Segmentation We used the method based on DDoP filter in [13] to segment the pulmonary fissure. The basic idea isWe toused match the the method local CT based image on DDoPusing a filter series in [of13 plate] to segment templates the with pulmonary different fissure. orientations. The basic When idea theis to lung match fissure the local surface CT image is detected, using a seriesthe matchi of plateng templates function with obtains different a high orientations. response Whenand the the correspondinglung fissure surface template is detected, direction the is matchingthe surface function normal obtainsdirection. a high Then, response in postprocessing, and the corresponding directional partitiontemplate directiondenoising isand the surfacethe fusion normal algorithm direction. based Then, on inthe postprocessing, lung fissure’s directionalfeatures, including partition consistencydenoising and of thethe fusion normal algorithm vector basedand onthe the largest lung fissure’sconnected features, area, includingare used consistency to remove of the the nonpulmonarynormal vector andfissure the surface largest connectedstructures area,and extract are used the to lung remove fissure the in nonpulmonary 3D space. The fissuresegmentation surface resultstructures is shown and extractin Figure the 2. lung fissure in 3D space. The segmentation result is shown in Figure2.

FigureFigure 2. 2. PulmonaryPulmonary fissure ﬁssure segmentation segmentation result. result.

Algorithms 2020, 13, 263 4 of 12

2.2.2. Lung Fissure Surface Classification Using a Point Cloud Region Growing Approach The overall process of classifying the lung fissures into the oblique fissure and horizontal fissure mainly includes two steps: (1) estimating the lung fissure surface’s normal vector and curvature based on point cloud data; (2) using the point cloud region growing approach to classify the lung fissure. The basic idea is to fuse the point cloud with similar properties (normal and curvature) in the neighborhood by the region growth algorithm, and then classify the pulmonary fissure according to the assumption that the oblique fissure area is greater than the horizontal fissure area. There are several advantages to representing the lung fissure as a point cloud, including (1) the point cloud data can better represent the geometry shape, and it is convenient to calculate the normal vector and curvature more accurately, (2) unordered point cloud data saves memory space and has high computing efficiency, and (3) mature point cloud processing tools and algorithms can be used to complete specific pulmonary fissure analysis tasks, such as pulmonary fissure classification. The specific classification process is as follows. Because the surface’s local area can be approximated as a plane, the problem of estimating the normal vector of a point on the surface can be solved by calculating the approximate plane’s normal vector. Estimating a plane by the least square method can be further simplified to analyze a covariance matrix formed for each neighborhood using principal component analysis [15]. The relationship between a covariance matrix and its eigenvalues and eigenvectors is formulated as Equation (1).

C v = λ v , i 0, 1, 2 , λ λ λ . (1) · i i · i ∈ { } 0 ≤ 1 ≤ 2 Each eigenvector of the covariance matrix shows a direction of a principal variation and the one associated with the smallest eigenvalue corresponds to the direction of the least variation, which is geometrically perpendicular to the direction of the most variation. This is the normal vector of the plane we simply estimate. In order to further estimate the surface curvature, the dispersion of the point cloud along the normal direction is selected as the measurement criterion. The total dispersion of the neighborhood point cloud, i.e., the sum of distances from all neighborhood points to their center, can be obtained by

Xk x c 2 = λ + λ + λ , (2) k i − k 0 1 2 i=1 where c denotes the center of gravity. Therefore, the proportion of the point cloud dispersion along the normal direction in total dispersion can be used to estimate surface curvature or variance as

λ c = 0 . (3) λ0 + λ1 + λ2

After obtaining the normal vector and curvature of the lung fissure surface, we could use them as a smooth constraint to realize the lung fissure clustering by the region growing method [16]. The algorithm is shown in Algorithm 1. The minimal curvature point is selected as the increasing seed point. If the normal vector intersection angle between the neighborhood point and the seed point is less than the set threshold, the neighborhood point is fused into the same category as the seed point and removed from the original point cloud set. If the curvature difference between the neighborhood point and the seed point is less than the set threshold, the neighborhood point is added to the seed point set for subsequent growth and the current seed point is removed from the seed point set. When the seed point set is empty, it indicates that this regional growing has been completed, and enters into the next region growing. This is repeated until all point clouds are marked. The angle threshold is set to 3◦ and the curvature threshold is set to 0.1. Figure3 shows an example result of point cloud clustering, where the larger area is the oblique fissure and the smaller is the horizontal fissure. Algorithms 2020, 13, x FOR PEER REVIEW 5 of 12

Algorithm 1 Regional growth clustering Algorithms 2020, 13, 263 5 of 12 Input: lung fissure point cloud 𝑋, surface normal vector 𝑁, surface curvature 𝑐, neighborhood

search function Ω, curvature threshold 𝑐, angle threshold 𝜃 AlgorithmOutput: 1regionalRegional clustering growth clustering set 𝑅 1: initialization:lung ﬁssure point 𝑅←∅, cloudavailable X, surface point normal cloud vector set:N , surface𝐴 ←𝑋 curvature; c, Input: neighborhood search function Ω, curvature threshold c , angle threshold θ 2: while 𝐴 𝑁𝑈𝐿𝐿 do th th Output: regional clustering set R 3: current growth region 𝑅 ←∅; 1: initialization: R ∅, available point cloud set: A X ; 4: current seed point← set 𝑆 ←∅; { } ← 2: while A , NULL do 5:3: 𝑃 current = minimum growth region curvatureRc point∅ ; in 𝐴; { } ← 4: current seed point set Sc ∅ ; 6: 𝑆 ← 𝑆 ∪𝑃, 𝑅{ ←} ←𝑅 ∪𝑃, 𝐴 ← 𝐴\𝑃; 5: Pmin = minimum curvature point in A; 7: for 𝑖=1 to size({𝑆}) do 6: Sc Sc Pmin, Rc Rc Pmin, A A Pmin ; { } ← { } ∪ { } ← { } ∪ { } ← { }\ 8:7: forneighborhoodi = 1 to size({ Spointc}) do cloud of current seed 𝐵 ←Ω𝑆 𝑖 ; 9:8: for neighborhood 𝑗 =0 to size({ point𝐵 cloud}) do of current seed Bc Ω(Sc i ) ; { } ← { } 10:9: forcurrentj = 0 to neighborhoodsize({Bc}) do point 𝑃 ←𝐵𝑗; 10: current𝐴 neighborhood𝑃 point𝑐𝑜𝑠 P𝑁𝑆j Bc j𝑖;, 𝑁𝑃 𝜃 11: if { } contain and ← then 1 n o 12:11: if {𝑅A} contain← 𝑅 P∪𝑃and; cos N Sc i , N P < θ then j − { { }} j th 13:12: 𝐴Rc ← 𝐴Rc\𝑃; P ; { } ← { } ∪ j 14:13: ifA 𝑐𝑃 𝑐A P ; then { }n←o{ }\ j 15:14: if c𝑆Pj ←< c𝑆ththen ∪𝑃; 15: Sc Sc P ; 16: end{ if} ← { } ∪ j 17:16: endend if if 18:17: endend for if end for 19:18: end for 19: end for 20: add the current growth region to the cluster set 𝑅 ← 𝑅 ∪ 𝑅; 20: add the current growth region to the cluster set R R Rc ; 21: end while { } ← { } ∪ { } 21: end while

(a) (b)

FigureFigure 3.3. ResultsResults of of pulmonary pulmonary fissure fissure classification: classification: (a) (originala) original pulmonary pulmonary fissure fissure point point cloud, cloud, and and(b) clustered (b) clustered pulmonary pulmonary fissure fissure point point cloud. cloud.

2.3.2.3. LungLung FissureFissure SurfaceSurface FittingFitting andand LungLung LobeLobe SegmentationSegmentation AsAs shownshown in in Figure Figure3 ,3, there there may may be be an an incomplete incomplete lung lung fissure fissure after after classification, classification, which which results results in thein the failure failure to segment to segment adjacent adjacent lobes directly.lobes directly. To overcome To overcome it, we used it, thewe methodused the named method multistage named B-splinemultistage fitting B-spline [17]. Onfitting the [17]. basis On of thethe existingbasis of lungthe existing fissure information,lung fissure weinformation, made up thewe missingmade up lung the fissuremissing and lung extended fissure itand to theextended lung boundary it to the tolung form boundary a complete to surfaceform a thatcomplete could surface segment that adjacent could lungsegment lobes. adjacent This method lung lobes. can take This into method account can both take the into calculation account both efficiency the calculation and the fitting efficiency accuracy. and Thethe fitting specific accuracy. mathematical The specific principle mathematical is as follows. principle is as follows. TheThe fittingfitting functionfunction f𝑓can can be be constructed constructed by by using using the the known known lung lung fissure fissure coordinate coordinate( x𝑥,𝑦,𝑧, y, z).. TheThe componentcomponent (𝑥,𝑦x, y) is is defineddefined asas thethe independent variable, while the the component component z𝑧 isis the the functionfunction value,value, thenthen thethe missingmissing points’points’ coordinatescoordinates can can be be obtained obtained by by the the fitting fitting function functionf (𝑓𝑥,x, y). . 𝑦 However, it is difficult to find a balance point between the shape smoothing and fitting accuracy in the B-spline fitting algorithm. The multistage B-spline fitting algorithm [17] can effectively overcome this problem by using top-down hierarchical control grid functions. Firstly, the coarse grid control Algorithms 2020, 13, x FOR PEER REVIEW 6 of 12

Algorithms 2020, 13, x FOR PEER REVIEW 6 of 12 However, it is difficult to find a balance point between the shape smoothing and fitting accuracy in theHowever, B-spline it fittingis difficult algorithm. to find aThe balance multistage point between B-spline the fitting shape algorithmsmoothing [17]and fittingcan effectively accuracy Algorithmsovercomein the B-spline2020 this, 13 problem, 263fitting byalgorithm. using top-down The multistage hierarchical B-spline control fitting grid functions. algorithm Firstly, [17] canthe coarseeffectively6 grid of 12 controlovercome function this problem was used by using to get top-down a rough hierarchicalestimate, and control then grid the functions.finer control Firstly, grid the function coarse gridwas iterativelycontrol function used to was fit the used previous to get estimationa rough estimate error, finally, and givingthen the a more finer ac controlcurate andgrid smoothfunction fitting was function was used to get a rough estimate, and then the finer control grid function was iteratively used function.iteratively used to fit the previous estimation error, finally giving a more accurate and smooth fitting to fit the previous estimation error, finally giving a more accurate and smooth fitting function. function.To verify the surface fitting effect, the incomplete right lung oblique fissure and horizontal To verify the surface fitting effect, the incomplete right lung oblique fissure and horizontal fissure fissureTo were verify taken the assurface the fitting fitting objects, effect, their the fittingincomplete results right are shownlung oblique in Figure fissure 4. It canand be horizontal seen that were taken as the fitting objects, their fitting results are shown in Figure4. It can be seen that the fitting thefissure fitting were surface taken was as the smooth fitting without objects, deviation, their fitting could results effectively are shown fill thein Figure missing 4. pulmonaryIt can be seen fissure that surface was smooth without deviation, could effectively fill the missing pulmonary fissure and ensure andthe fitting ensure surface that it was smoothconsistent without with the deviation, existing could pulmonary effectively fissure. fill the missing pulmonary fissure that it was consistent with the existing pulmonary fissure. and ensure that it was consistent with the existing pulmonary fissure.

(a) (b) (c) (d)

Figure (4.a )Results of lung fissure surface(b) fitting: (a) original right(c )lung oblique fissure, (b) fitted(d) right lungFigure oblique 4. Results fissure, of lung (c) originalfissure fissure surface right lung fitting:fitting: horizontal ( a) original fissure, right andlung ( doblique ) fitted fissure, fissure,right lu ( ngb) fittedfittedhorizontal right fissurelung obliqueoblique. fissure, fissure, (c ()c original) original right right lung lung horizontal horizontal fissure, fissure, and (andd) fitted (d) rightfitted lung right horizontal lung horizontal fissure. fissure. Through thethe aboveabove steps,steps, the the complete complete oblique oblique fissure fissure and and horizontal horizontal fissure fissure were were fitted, fitted, and and the lungthe lungThrough could could be the easilybe easilyabove separated separated steps, the into into complete independent independent oblique lung lung fissure lobe lobe unitsand units horizontal by by using using these thesefissure surfaces. surfaces. were fitted, SinceSince and the rightthe lung lung could contains be easily three separated lobes, we into divided independent it in twotwo lung steps.steps. lobe Firstly, Firstly, units by the using obliqueoblique these fissurefissure surfaces. surface Since was the usedright to lung segment contains the threeright lobes,lung into we upper divided and it lower in two parts, parts, steps. where where Firstly, the the the lower lower oblique half half wasfissure was the the surface lower lower lobe. lobe.was Thenused to wewe segment usedused thethe the horizontal horizontal right lung fissure fissure into surfaceupper surface and to to segment lower segment parts, the the upper where upper half the half into lower into upper halfupper lobe was lobe and the middleandlower middle lobe. lobe. Thelobe.Then segmentation Thewe usedsegmentation the resultshorizontal results are shownfissure are shown insurface Figure in Figureto5. segment 5. the upper half into upper lobe and middle lobe. The segmentation results are shown in Figure 5.

(a) (b) (c) (d)

Figure 5. Results of(a lung) lobe segmentation:(b) ( a) original right(c) lung, (b) segmentationsegmentation(d) using oblique fissure,fissure,Figure 5. (c )Results segmentation of lung usingusing lobe segmentation: horizontalhorizontal fissure,fissure, (a) original andand ((dd)) segmentationrightsegmentation lung, (b) result.result segmentation. using oblique fissure, (c) segmentation using horizontal fissure, and (d) segmentation result. 3. Experimental Results 3. ExperimentalWe used data Results from 55 CT scans to verify verify the the segmentation segmentation performance performance of of the the proposed proposed method. method. This dataset was provided by the organizers of LOLA11 (Lobe and Lung Analysis 2011 Competition). This datasetWe used was data provided from 55 byCT the scans organizers to verify of the LO segmentationLA11 (Lobe and performance Lung Analysis of the 2011 proposed Competition). method. It was obtained from a number of clinical institutions using different CT scanning equipment and ItThis was dataset obtained was fromprovided a number by the oforganizers clinical institutions of LOLA11 using(Lobe differentand Lung CT Analysis scanning 2011 equipment Competition). and covered different degrees of pathological data. At the same time, the competition organizers also coveredIt was obtained different from degrees a number of pathological of clinical data.institutions At the usingsame differenttime, the CTcompetition scanning equipmentorganizers alsoand provided the manual segmentation results for the performance evaluation of the segmentation methods. providedcovered different the manual degrees segmentation of pathological results data. for Atth ethe performance same time, evaluationthe competition of the organizers segmentation also However, the gold standard data will not be released to the public and the segmentation results can methods.provided However,the manual the segmentation gold standard results data will for nothte beperformance released to theevaluation public and of thethe segmentation only be uploaded to the competition website [18] for online evaluation. resultsmethods. can However, only be uploaded the gold tostandard the competition data will websitenot be released [18] for onlineto the evaluation.public and the segmentation In this section, the proposed method’s performance was verified from two aspects, including resultsIn canthis only section, be uploaded the proposed to the method’s competition perfor websitemance [18] was for verified online evaluation.from two aspects, including visual and quantitative evaluation. Mevislab [19] was used for 3D visualization. The evaluation score visualIn and this quantitative section, the evaluation. proposed method’sMevislab [19]perfor wasmance used wasfor 3D verified visualization. from two The aspects, evaluation including score was obtained by observing and quantifying the size of the coincidence area between the segmentation visual and quantitative evaluation. Mevislab [19] was used for 3D visualization. The evaluation score results and the real lung lobes. The method for evaluating the size is

Lseg Lgt OV = ∩ (4) Lseg Lgt ∪ Algorithms 2020, 13, x FOR PEER REVIEW 7 of 12 was obtained by observing and quantifying the size of the coincidence area between the segmentation results and the real lung lobes. The method for evaluating the size is Algorithms 2020, 13, 263 7 of 12 𝐿 ∩𝐿 𝑂𝑉 = (4) 𝐿 ∪𝐿 where L is the lung segmentation result and L is the lung lobe gold standard. where 𝐿seg is the lung segmentation result and 𝐿gt is the lung lobe gold standard.

3.1.3.1. Visual Visual Evaluation Evaluation ThisThis section section shows shows the the representative representative lung lung lobe lobe segmentation segmentation results, results, and and compares compares our our method method withwith the the segmentation segmentation method method based based on on airway airway or or vessel vessel anatomical anatomical information. information. AsAs shown shown in in Figure Figure 6,6, five five representative representative lung lung lobe lobe segmentation segmentation examples examples were were selected selected here. here. RowsRows 1 1 to to 5 5 are are case case 05, 05, case case 07, 07, case case 11, 11, case case 13 13 and and case case 42 42 in in LOLA11, LOLA11, respectively. respectively. Columns Columns 1 1 to to 4 4 showshow the the pulmonary pulmonary lobe lobe segmentation 3D3D resultsresults andand thethe e effectsffects of of superimposing superimposing the the original original data data on onthe the segmentation segmentation results’ results’ transverse, transverse, coronal coronal and and sagittal sagittal planes, planes, respectively. respectively. It canIt can be be seen seen from from the theL05 L05 that that although although case case 05 showed05 showed severe severe deformation deformation due due to scoliosis, to scoliosis, the segmentation the segmentation results results using usingthe proposed the proposed method method were not were affected not affected and the boundariesand the boundaries of the lung of lobe the segmentationlung lobe segmentation results were resultsclose to were the realclose boundaries. to the real Althoughboundaries. case Although 07 was a low-dosecase 07 was CT a image low-dose and theCT noiseimage interference and the noise was interferenceserious, the was segmentation serious, the results segmentation for L07 show results that for this L07 method show can that still this obtain method a better can still segmentation obtain a bettereffect. segmentation Pulmonary fibrosis effect. wasPulmonary serious infibrosis case 11, was but se becauserious in ofcase the 11, strong but because noise suppression of the strong ability noise of suppressionthe detection ability method, of the the detection lung lobe method, boundaries the lung in L11 lobe were boundaries close to the in realL11 pulmonarywere close to fissure the real and pulmonarywere not a fffissureected by and noise. were Case not 42, affected from an by emphysema noise. Case patient, 42, from also an achieved emphysema better segmentationpatient, also achievedresults, asbetter shown segmentation in L42. Experimental results, as results shown showed in L42. that Experimental the proposed results segmentation showed method that the not proposedonly achieved segmentation good segmentation method not performance only achieved on good high-quality segmentation CT data, performance but also could on high-quality segment the CTlung data, lobe but correctly also could in pathological segment the cases. lung lobe correctly in pathological cases.

FigureFigure 6. 6. RepresentativeRepresentative results results of of lo lobebe segmentation segmentation from from LOLA11 LOLA11..

Compared with the lung lobe segmentation algorithm proposed by Doel et al. in [12], our algorithm has better performance in detecting visible pulmonary fissures because it is not affected by the Algorithms 2020, 13, x FOR PEER REVIEW 8 of 12 Algorithms 2020, 13, x FOR PEER REVIEW 8 of 12 AlgorithmsCompared2020, 13, 263with the lung lobe segmentation algorithm proposed by Doel et al. in [12],8 ofour 12 algorithmCompared has better with performance the lung lobe in detectingsegmentation visible algorithm pulmonary proposed fissures becauseby Doel itet is al.not in affected [12], our by thealgorithm segmentation has better accuracy performance of airways in detecting and vessels. visible The pulmonary first and fissures second becauselines in itFigure is not 7affected show the by segmentationrepresentativethe segmentation accuracyeffects accuracy of superimposing of airwaysof airways and and the vessels. originalvessels. The Thedata first firston andthe and segmentation second second lines lines results’ inin FigureFigure coronal7 7show show planes the the representativeusing Doel’s method eeffectsffects and ofof superimposing ours, respectively. the original From left data to onright the are segmentation the results of results’ case 19, coronal case 26, planes case using36, and Doel’s Doel’s case method53. method For Doel’sand and ours, ours, method, respectively. respectively. the areas From Frommarked left left to with right to righta arered are theellipse theresults resultswere of casethe of location case19, case 19, of case26, error case 26, casesegmentation.36, and 36, andcase case53. The For 53. possible ForDoel’s Doel’s reasonsmethod, method, for the these theareas areas errors marked marked included with with a two red a red aspects:ellipse ellipse were(1) were the the the pulmonary location location of offissure error segmentation.candidate areas The The located possible by thereasons airway for andthese vessel errors information included two were aspects: not accurate, (1) the pulmonary and (2) the fissurefissure weak candidatepulmonary areas fissures located with by uneven the airway gray scale and were vessel missed. information were not accurate, and (2) the weak pulmonary fissuresfissures withwith unevenuneven graygray scalescale werewere missed.missed.

Figure 7. Comparison between present method (second line) and Doel’s method (first line) when the Figure 7. Comparison between present method (second line) and Doel’s method (first line) when the pulmonaryFigure 7. Comparison fissure is visible between. present method (second line) and Doel’s method (first line) when the pulmonary fissure is visible. pulmonary fissure is visible. Besides, segmentation results for cases with moderate and low scores are shown in Figure 8. For Besides, segmentation results for cases with moderate and low scores are shown in Figure8. case Besides,27, the rightsegmentation middle loberesults from for casesour method with moder andate Doel’s and lowmethod scores was are inaccurateshown in Figure due to 8. Forthe For case 27, the right middle lobe from our method and Doel’s method was inaccurate due to the incompletecase 27, the horizontal right middle fissure. lobe For from case our 31, methodcase 44 andand caseDoel’s 45, methodthe lung was segmentation inaccurate results due to were the incomplete horizontal fissure. For case 31, case 44 and case 45, the lung segmentation results were poorincomplete because horizontal of the invisible fissure. lungFor caseboundaries 31, case caused44 and bycase lung 45, disease,the lung which segmentation directly resultsaffected were the poor because of the invisible lung boundaries caused by lung disease, which directly affected the performancepoor because of of the the lobe invisible segmentati lung on.boundaries Furthermore, caused as bydemonstrated lung disease, in L45,which Doel’s directly method affected had the performance of the lobe segmentation. Furthermore, as demonstrated in L45, Doel’s method had abilityperformance to construct of the lobethe rightsegmentati middleon. lobe Furthermore, for the CT as image demonstrated with invisible in L45, lung Doel’s fissures, method while had the the ability to construct the right middle lobe for the CT image with invisible lung fissures, while the proposedability to constructmethod could the rightnot deal middle with lobe this situation.for the CT image with invisible lung fissures, while the proposed method could not deal with this situation. proposed method could not deal with this situation.

L27 L31 L44 L45

Figure 8. ComparisonL27 between our our method L31 (the second ro row)w)L44 and Doel’s method L45(the first ﬁrst row) for moderateFigure 8. Comparison and low score between segmentation our method results. (the second row) and Doel’s method (the first row) for moderate and low score segmentation results. Algorithms 2020, 13, 263 9 of 12

3.2. Quantitative Evaluation The quantitative evaluation results came from the online evaluation system of the competition website. The 55 LOLA11 cases’ segmentation results were uploaded to the competition website, and the background evaluation system executed the evaluation procedure and fed back the evaluation results. Specifically, according to Formula (4), each lung lobe’s coincidence degree in each data object was calculated, and then the coincidence degree mean value from each type of lung lobe was calculated. The overall score was the average value of the five lung lobes’ mean coincidence degrees. Furthermore, we used six other quantitative evaluation indicators based on the mean statistics including variance (SD), minimum (Min), quarter (Q1), median (Median), three-quarter (Q3) and maximum (Max). Table1 presents the total quantitative evaluation results. The proposed method’s overall score was 0.84. The highest and lowest scores of single class assessment were from the left upper lobe (0.93) and right middle lobe (0.65), respectively. The variance of the right middle lobe was 0.4, which indicated that the data difference had a great influence on the segmentation performance and the reason was that there were many incomplete right lung horizontal fissures. For a more comprehensive understanding, including the effect of each case’s segmentation in the LOLA11 dataset, Figure9 describes the specific coincidence degree score of each case.

Table 1. Overall quantitative assessment of lobe segmentation on LOLA11 dataset.

Mean SD Min Q1 Median Q3 Max Left lower lobe 0.89 0.23 0.0 0.95 0.97 0.98 0.99 Left upper lobe 0.93 0.18 0.02 0.96 0.99 0.99 1 Right upper lobe 0.84 0.27 0.0 0.8 0.96 0.99 1 Right middle lobe 0.65 0.4 0.0 0.06 0.82 0.96 1 Right lower lobe 0.88 0.24 0.0 0.93 0.97 0.98 0.99 Total score 0.84

As shown in Figure9, many wrong segmentation results exist in the right middle lobes and there is still a lot of room for improvement. The reason for the poor performance was missing the detection of the horizontal fissure because it was difficult to image the horizontal fissures in the CT images when the horizontal fissures were nearly parallel to the CT scanning plane. To further verify the performance of the proposed method, we compared the average score from each lung lobe between our method and two representative AI-based methods (a sequence of convolutional neural networks for marginal learning [20] and progressive dense V-Network [4]). The quantitative evaluation results on each lung lobe are shown in Table2 (LobeNet_V2 is the name of the net in [20] given by authors for LOLA11). It indicates that our proposed method achieved a similar performance for most lung lobes except the right middle lobe, and there is even better segmentation performance on the left lung compared with V-net. Moreover, the proposed algorithm does not need a large number of labeled data for training and has advantages in operating efficiency.

Table 2. Comparison between our proposed method and AI-based methods.

LobeNet_V2 V-Net The Proposed Method Left lower lobe 0.91 0.88 0.89 Left upper lobe 0.95 0.92 0.93 Right lower lobe 0.96 0.92 0.88 Right middle lobe 0.87 0.76 0.65 Right upper lobe 0.94 0.91 0.84 Algorithms 2020, 13, x FOR PEER REVIEW 9 of 12

3.2. Quantitative Evaluation The quantitative evaluation results came from the online evaluation system of the competition website. The 55 LOLA11 cases’ segmentation results were uploaded to the competition website, and the background evaluation system executed the evaluation procedure and fed back the evaluation results. Specifically, according to Formula (4), each lung lobe’s coincidence degree in each data object was calculated, and then the coincidence degree mean value from each type of lung lobe was calculated. The overall score was the average value of the five lung lobes’ mean coincidence degrees. Furthermore, we used six other quantitative evaluation indicators based on the mean statistics including variance (SD), minimum (Min), quarter (Q1), median (Median), three-quarter (Q3) and maximum (Max). Table 1 presents the total quantitative evaluation results. The proposed method’s overall score was 0.84. The highest and lowest scores of single class assessment were from the left upper lobe (0.93) and right middle lobe (0.65), respectively. The variance of the right middle lobe was 0.4, which indicated that the data difference had a great influence on the segmentation performance and the reason was that there were many incomplete right lung horizontal fissures. For a more comprehensive understanding, including the effect of each case’s segmentation in the LOLA11 dataset, Figure 9 describes the specific coincidence degree score of each case.

Table 1. Overall quantitative assessment of lobe segmentation on LOLA11 dataset.

As shown in Figure 9, many wrong segmentation results exist in the right middle lobes and there is still a lot of room for improvement. The reason for the poor performance was missing the detection of the horizontal fissure because it was difficult to image the horizontal fissures in the CT images when the horizontal fissures were nearly parallel to the CT scanning plane. To further verify the performance of the proposed method, we compared the average score from each lung lobe between our method and two representative AI-based methods (a sequence of convolutional neural networks for marginal learning [20] and progressive dense V-Network [4]). The quantitative evaluation results on each lung lobe are shown in Table 2 (LobeNet_V2 is the name of the net in [20] given by authors for LOLA11). It indicates that our proposed method achieved a similar performance for most lung lobes except the right middle lobe, and there is even better segmentation performance on the left lung compared with V-net. Moreover, the proposed algorithm does not need Algorithmsa large number2020, 13, 263of labeled data for training and has advantages in operating efficiency. 10 of 12

LOLA11：01−30

0.8

0.6

0.4

0.2

0 L01 L02 L03 L04 L05 L06 L07 L08 L09 L10 L11 L12 L13 L14 L15 L16 L17 L18 L19 L20 L21 L22 L23 L24 L25 L26 L27 L28 L29 L30

Algorithms 2020, 13, x FOR PEER REVIEW lll lul rll rml rul 10 of 12

(a)

LOLA11：31−55

0.8

0.6

0.4

0.2

0 L31 L32 L33 L34 L35 L36 L37 L38 L39 L40 L41 L42 L43 L44 L45 L46 L47 L48 L49 L50 L51 L52 L53 L54 L55

lll lul rll rml rul

(b)

FigureFigure 9.9. QuantitativeQuantitative evaluationevaluation resultsresults ofof lunglung lobelobe segmentationsegmentation in in the the LOLA11 LOLA11 dataset: dataset: ((aa)) resultsresults ofof casecase 0101−case 30, and and ( (bb)) results results of of case case 31 31−casecase 55. 55. The The lll, lll, lul, lul, rll, rll, rml rml and and rul rul are are left left lower lower lobe, lobe, left − − leftupper upper lobe, lobe, right right lower lower lobe, lobe, right right middle middle lobe lobe and and right right upper upper lobe, lobe, respectively. respectively.

4. Discussion Table 2. Comparison between our proposed method and AI-based methods.

The proposedlung lobe segmentation LobeNet_V2 framework V-Net has the following The Proposed characteristics: Method (1) it does not rely on anatomicalLeft lower lobe knowledge such 0.91 as airways or vessels, 0.88 but only relies on 0.89 the assumption that the lung fissureLeft canupper be representedlobe as a0.95 continuous surface, 0.92 (2) a fast and highly 0.93 e fficient multistage B-spline surface-fittingRight lower lobe algorithm is used 0.96 to fill the lung 0.92 fissure hole and extend 0.88 the lung fissure to the lung boundaryRight middle to form lobe a complete 0.87 lung fissure surface, 0.76 and (3) according 0.65 to the difference in the normal vectorRight and upper curvature lobe in different 0.94 pulmonary fissures, 0.91 the right pulmonary 0.84 fissure is classified as the horizontal and oblique fissure. 4. DiscussionA large number of researchers use prior knowledge of lung anatomical structure to locate the lung fissure area, so as to reduce the false positive detection and estimate the invisible pulmonary fissure.The Although proposed this lung type lobe of method segmentation can improve framework the detection has the e ffifollowingciency, it characteristics: is limited by the (1) accuracy it does ofnot prior rely anatomical on anatomical structure knowledge knowledge. such as This airways limitation or vessels, can be overcomebut only relies by directly on the using assumption the density that distributionthe lung fissure and shapecan be features represented of the as pulmonary a continuous fissure surface, in 3D (2) images, a fast andand betterhighly detection efficient resultsmultistage can beB-spline obtained surface-fitting in lung data algorithm with visible is used lung fissures,to fill the such lung as fissure the experimental hole and ex resultstend the of lung theproposed fissure to method.the lung Inboundary addition, to comparedform a complete with the lung deep fissure learning surface, method, and (3) which according largely to dependsthe difference on a largein the amountnormal ofvector labeled and data curvature and uniformity in different of pulmonary image acquisition fissures, devices, the right the pulmonary proposed method fissure is has classified higher economyas the horizontal and applicability. and oblique fissure. ExperimentalA large number results of researchers show that this use framework prior knowledg givese a of good lung segmentation anatomical structure performance to locate for most the CTlung images, fissure but area, for so some as to others reduce the the horizontal false posi fissuretive detection may be easily and missedestimate during the invisible pulmonary pulmonary fissure detection.fissure. Although This is because this type the of horizontal method can fissure improve often appearsthe detection invisible efficiency, in CT images. it is limited To solve by this the problem,accuracy itof is prior necessary anatomical to integrate structure the anatomicalknowledge. information This limitation into can the lungbe overcome fissure segmentation by directly using and surfacethe density fitting distribution stages, that and is, to shape use the features global informationof the pulmonary to extract fissure the discontinuousin 3D images, pulmonary and better fissuredetection surface, results which can willbe obtained be a future in job.lung Furthermore, data with visible for the lung deep fissures, learning such method, as the we experimental will explore results of the proposed method. In addition, compared with the deep learning method, which largely depends on a large amount of labeled data and uniformity of image acquisition devices, the proposed method has higher economy and applicability. Experimental results show that this framework gives a good segmentation performance for most CT images, but for some others the horizontal fissure may be easily missed during pulmonary fissure detection. This is because the horizontal fissure often appears invisible in CT images. To solve this problem, it is necessary to integrate the anatomical information into the lung fissure segmentation and surface fitting stages, that is, to use the global information to extract the discontinuous pulmonary fissure surface, which will be a future job. Furthermore, for the deep learning method, we will explore integration of prior knowledge, such as global geometry or traditional image processing knowledge, into a convolution network, and hope to reduce the dependence on labeled data. Algorithms 2020, 13, 263 11 of 12 integration of prior knowledge, such as global geometry or traditional image processing knowledge, into a convolution network, and hope to reduce the dependence on labeled data.

Author Contributions: Conceptualization, H.Z.; format analysis, H.Z.; investigation, X.C.; methodology, X.C. and H.Z.; project administration, P.Z.; resources, P.Z.; software, X.C.; validation, P.Z.; writing—original draft, X.C.; writing—review & editing, X.C. and H.Z. All authors have read and agreed to the published version of the manuscript. Funding: This research was funded by National Natural Science Foundation of China, grant number 61571184 and 61733004, and funded the COVID-19 special item in Changsha Science and Technology Plan, grant number kq2001014. Acknowledgments: We acknowledge the organizers of LOLA11 for providing the dataset and algorithm performance evaluation system. Conﬂicts of Interest: The authors declare no conﬂict of interest.

References

1. Han, M.K.; Agusti, A.; Calverley, P.M.; Celli, B.R.; Criner, G.; Curtis, J.L.; Fabbri, L.M.; Goldin, J.G.; Jones, P.W.; MacNee, W.; et al. Chronic obstructive pulmonary disease phenotypes: The future of COPD. Am. J. Respir. Crit. Care Med. 2010, 182, 598–604. [CrossRef][PubMed] 2. Doel, T.; Gavaghan, D.J.; Grau, V. Review of automatic pulmonary lobe segmentation methods from CT. Comput. Med. Imaging Graph. 2015, 40, 13–29. [CrossRef][PubMed] 3. George, K.; Harrison, A.P.; Jin, D.; Xu, Z.; Mollura, D.J. Pathological pulmonary lobe segmentation from CT images using progressive holistically nested neural networks and random walker. In Proceedings of the Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, Quebec, QC, Canada, 14 September 2017; Springer: Berlin/Heidelberg, Germany, 2017; pp. 195–203. 4. Imran, A.A.Z.; Hatamizadeh, A.; Ananth, S.P.; Ding, X.; Terzopoulos, D.; Tajbakhsh, N. Automatic segmentation of pulmonary lobes using a progressive dense V-Network. arXiv 2019, arXiv:1902.06362. 5. Ferreira, F.T.; Sousa, P.; Galdran, A.; Sousa, M.R.; Campilho, A. End-to-end supervised lung lobe segmentation. In Proceedings of the International Joint Conference on Neural Networks (IJCNN), Rio De Janeiro, Brazil, 8–13 July 2018; IEEE: Piscataway, NJ, USA, 2018; pp. 1–8. 6. Ukil, S.; Reinhardt, J.M. Smoothing lung segmentation surfaces in three-dimensional x-ray ct images using anatomic guidance1. Acad. Radiol. 2005, 12, 1502–1511. [CrossRef][PubMed] 7. Sun, X.; Zhang, H.; Duan, H. 3D computerized segmentation of lung volume with computed tomography. Acad. Radiol. 2006, 13, 670–677. [CrossRef][PubMed] 8. Pu, J.; Roos, J.; Chin, A.Y.; Napel, S.; Rubin, G.D.; Paik, D.S. Adaptive border marching algorithm: Automatic lung segmentation on chest CT images. Comput. Med. Imaging Graph. 2008, 32, 452–462. [CrossRef][PubMed] 9. Saita, S.; Kubo, M.; Kawata, Y.; Niki, N.; Ohmatsu, H.; Moriyama, N. An algorithm for the extraction of pulmonary fissures from low-dose multislice CT image. Syst. Comput. Jpn. 2006, 37, 63–76. [CrossRef] 10. Pu, J.; Leader, J.K.; Zheng, B.; Knollmann, F.; Fuhrman, C.; Sciurba, F.C.; Gur, D. A computational geometry approach to automated pulmonary fissure segmentation in CT examinations. IEEE Trans. Med. Imaging 2009, 28, 710. [PubMed] 11. Ukil, S.; Reinhardt, J.M. Anatomy-guided lung lobe segmentation in X-ray CT images. IEEE Trans. Med. Imaging 2008, 28, 202–214. [CrossRef][PubMed] 12. Doel, T.; Matin, T.N.; Gleeson, F.V.; Gavaghan, D.J.; Grau, V. Pulmonary lobe segmentation from CT images using fissureness, airways, vessels and multilevel B-splines. In Proceedings of the Biomedical Imaging (ISBI), 2012 9th IEEE International Symposium on, Barcelona, Spain, 2–4 May 2012; IEEE: Piscataway, NJ, USA, 2012; pp. 1491–1494. 13. Zhao, H.; Stoel, B.C.; Staring, M.; Bakker, M.E.; Stolk, J.; Zhou, P.; Xiao, C. A framework for pulmonary fissure segmentation in 3D CT images using a directional derivative of plate filter. Signal Process. 2020, 173, 107602. [CrossRef] 14. VanRikxoort, E.M.; deHoop, B.; Viergever, M.A.; Prokop, M.; Van Ginneken, B.B. Automatic lung segmentation from thoracic computed tomography scans using a hybrid approach with error detection. Med. Phys. 2009, 36, 2934–2947. [CrossRef][PubMed] Algorithms 2020, 13, 263 12 of 12

15. Wold, S.; Esbensen, K.; Geladi, P. Principal component analysis. Chemom. Intell. Lab. Syst. 1987, 2, 37–52. [CrossRef] 16. Region Growing Approach. Available online: http://pointclouds.org/documentation/tutorials (accessed on 14 October 2020). 17. Lee, S.; Wolberg, G.; Shin, S.Y. Scattered data interpolation with multilevel B-splines. IEEE Trans. Vis. Comput. Graph. 1997, 3, 228–244. [CrossRef] 18. Competition Website. Available online: https://lola11.grand-challenge.org (accessed on 14 October 2020). 19. MeVisLab. Available online: http://www.mevislab.de (accessed on 14 October 2020). 20. Gerard, S.E.; Reinhardt, J.M. Pulmonary lobe segmentation using a sequence of convolutional neural networks for marginal learning. In Proceedings of the Biomedical Imaging (ISBI), 2019 16th IEEE International Symposium on Biomedical Imaging, Venezia, Italy, 24–27 April 2019; IEEE: Piscataway, NJ, USA, 2019; pp. 1207–1211.

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional aﬃliations.

© 2020 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).