DOCSLIB.ORG

Chuan Thesis Draft Final

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Chuan Thesis Draft Final Prediction of Electron Cutout Factors Using Residual Neural Network by Chuan He Graduate Program in Medical Physics Duke University Date:_______________________ Approved: ___________________________ Qiuwen Wu, Supervisor ___________________________ Jun Lian ___________________________ Shiva Das ___________________________ Yvonne Mowery Thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in the Graduate Program in Medical Physics in the Graduate School of Duke University 2019 iv ABSTRACT Prediction of Electron Cutout Factors Using Residual Neural Network by Chuan He Graduate Program in Medical Physics Duke University Date:_______________________ Approved: ___________________________ Qiuwen Wu, Supervisor ___________________________ Jun Lian ___________________________ Shiva Das ___________________________ Yvonne Mowery An abstract of a thesis submitted in partial fulfillment of the requirements for the degree of Master of Science in the Graduate Program in Medical Physics in the Graduate School of Duke University 2019 iv Copyright by Chuan He 2019 Abstract Background: Monte Carlo and square root are two commonly used calculation methods to predict electron cutout factors in clinic. Monte Carlo is accurate but time consuming, while square root is faster but less accurate for cutouts with highly irregular shapes. Purpose: Simplify and develop an efficient residual neural network model to predict electron cutout factors accurately and instantly. Methods: 281 clinical cutouts were screened, and 12 groups were designed combining four different electron energies (6, 9, 12 and 15 MeV) and three different cones (10, 14 and 20 cm). The cutout factors of 35 previously used electron cutouts were calculated by Monte Carlo simulation and also measured with a solid water phantom and an ion chamber for validation of Monte Carlo accuracy. To solve the issue of sparse training data, 600 cutout samples were created for each group based on modifications of previously screened clinical cutouts. Cutout factors of these 600 samples were calculated with Monte Carlo simulation. 400 samples were randomly selected as training data, 50 as validating and the remaining 150 as testing. 1D Distance histograms were calculated as model input instead of 2D images to accelerate the training process. 1D Residual neural network with four residual blocks and three linear blocks was used. Performance of the trained model was evaluated with testing data, and the accuracy of the model was iv compared with square root method for eight selected cutouts with highly irregular shapes. Results: The Monte Carlo calculated cutout factor agreed with the measurement within 0.7±0.5% on average. During the training process, the model started to converge within 20 epochs with 30 seconds. For model prediction, mean errors and maximum discrepancies for each energy and cone combinations were all within 1%, and the average overall error was 0.2±0.16%. Compared to square root method, the trained model performed better for cutouts with highly irregular shapes. Conclusion: An efficient residual neural network model was simplified and developed, which is capable of estimating electron cutout factors accurately and instantly. v Contents Abstract .......................................................................................................................................... iv List of Tables ................................................................................................................................. ix List of Figures ................................................................................................................................ x Acknowledgements .................................................................................................................... xii 1. Introduction ............................................................................................................................... 1 1.1 Electron beam radiation therapy .................................................................................... 2 1.1.1 Ionizing radiation ........................................................................................................ 2 1.1.2 External beam radiation therapy ............................................................................... 3 1.1.3 Dose calculation for electron beams ......................................................................... 5 1.1.4 Prediction of electron output factors in the literature ............................................ 7 1.1.4.1 Equivalent square and square root .................................................................... 8 1.1.4.2 Sector integration and CMCI ............................................................................ 10 1.1.4.3 Pencil beam ......................................................................................................... 11 1.1.4.4 Monte Carlo Simulation .................................................................................... 12 1.2 Deep learning .................................................................................................................. 13 1.2.1 Regression and classification ................................................................................... 14 1.2.2 Overfitting .................................................................................................................. 15 1.2.3 Convolutional neural network ................................................................................ 17 1.2.4 Residual neural network .......................................................................................... 19 1.2.5 Application of deep learning in medicine ............................................................. 20 vi 1.3 Summary .......................................................................................................................... 22 1.4 Project Objective ............................................................................................................. 23 2. Methods .................................................................................................................................... 24 2.1 Preparations of cutouts and cutout factors ................................................................. 24 2.1.1 Group design .............................................................................................................. 24 2.1.2 Create plans for each case in RayStation ................................................................ 27 2.1.3 Calculation of cutout factors using Monte Carlo method ................................... 30 2.2 Model development ....................................................................................................... 31 2.2.1 Distance histogram for cutouts ............................................................................... 31 2.2.2 ResNet model structure ............................................................................................ 33 2.3 Model training ................................................................................................................ 36 2.4 Model performance evaluation .................................................................................... 38 3. Results ....................................................................................................................................... 39 3.1 Monte Carlo accuracy .................................................................................................... 39 3.2 Accuracy of trained models .......................................................................................... 40 3.3 Comparisons of square root and ResNet model prediction accuracy .................... 45 4. Discussion ................................................................................................................................ 47 4.1 Is Monte Carlo simulation accurate enough? ............................................................. 47 4.2 Advantages and disadvantages of using distance histogram .................................. 47 4.3 Model performance ........................................................................................................ 49 4.4 Is deep learning a better choice compared to square root? ...................................... 50 4.5 Future work ..................................................................................................................... 51 vii 5. Conclusion ................................................................................................................................ 53 References ..................................................................................................................................... 54 viii List of Tables Table 1: Screened clinical cases for electron beam radiation therapy ................................. 25 Table 2: Summary of group design .......................................................................................... 27 Table 3: Hyperparameters used for model training. .............................................................. 37 Table 4: Measured (MS) and calculated (MC) cutout factors for eight clinical cutouts. ... 40 Table 5: Prediction errors for trained models. ........................................................................ 41 Table 6: Errors for cutout factor estimation using square root and ResNet model. (MC for Monte Carlo simulation, ML for model prediction, and
Load more

Recommended publications
  • Localized Motion Artifact Reduction on Brain MRI Using Deep Learning with Effective Data Augmentation Techniques
    Localized Motion Artifact Reduction on Brain MRI Using Deep Learning with Effective Data Augmentation Techniques Yijun Zhaoa , Jacek Ossowskib , Xuming Wanga , Shangjin Lia , Orrin Devinskyc , Samantha P. Martinc , Heath R. Pardoec aFordham University, 113 60th Street, New York, NY 10023, USA bQS Investors, LLC, New York, NY 10022, USA cGrossman Medical School, 145 East 32nd Street, New York, NY 100162, USA Abstract In-scanner motion degrades the quality of magnetic resonance imaging (MRI) thereby reducing its utility in the detection of clinically relevant abnormalities. We introduce a deep learning-based MRI artifact reduction model (DMAR) to localize and correct head motion artifacts in brain MRI scans. Our approach integrates the latest advances in object detection and noise reduction in Computer Vision. Specifically, DMAR employs a two-stage approach: in the first, degraded regions are detected using the Single Shot Multibox Detector (SSD), and in the second, the artifacts within the found regions are reduced using a convolutional autoencoder (CAE). We further introduce a set of novel data augmentation techniques to address the high dimensionality of MRI images and the scarcity of available data. As a result, our model was trained on a large synthetic dataset of 225,000 images generated from 375 whole brain T1-weighted MRI scans. DMAR visibly reduces image artifacts when applied to both synthetic test images and 55 real-world motion-affected slices from 18 subjects from the multi-center Autism Brain Imaging Data Exchange (ABIDE) study. Quantitatively, depending on the level of degradation, our model achieves a 27.8%–48.1% reduction in RMSE and a 2.88–5.79 dB gain in PSNR on a 5000-sample set of synthetic images.
    [Show full text]
  • Rheinische Friedrich-Wilhelms-Universität Bonn
    Rheinische Friedrich-Wilhelms-Universitat¨ Bonn Master thesis Segmentation of Plant Root MRI Images Author: First Examiner: Ali Oguz Uzman Prof. Dr. Sven Behnke Second Examiner: Prof. Dr. Thomas Schultz Advisor: Prof. Dr. Sven Behnke Submitted: 24.10.2018 Declaration of Authorship I declare that the work presented here is original and the result of my own investigations. Formulations and ideas taken from other sources are cited as such. It has not been submitted, either in part or whole, for a degree at this or any other university. Location, Date Signature Abstract The plant roots have been a long standing research interest due to their crucial role for plants. As a non-invasive method, Magnetic Resonance Imaging (MRI) is used to overcome the opaque nature of soil and obtain 3D visualizations of plant roots. Existing algorithms fail to extract the structural model of the root when the environment (soil) is noisy and the resolution of MRI images is low. To this end, we develop a convolutional neural network to segment plant root MRI images as root vs non-root. The resulting segmentations have a higher resolution than their input MRI data. Our convolutional neural network is based on RefineNet, a state of the art se- mantic segmentation method. As pretrained networks used in RefineNet expect 2D images, we use PCA to reduce 3D data into 2D RGB images for feature ex- traction. We test different loss functions to overcome the class imbalance problem between root and non-root voxels. The provided data is insufficient for training a neural network. Thus, we develop data augmentation processes to create synthetic training data.
    [Show full text]
  • Data and Image Prior Integration for Image Reconstruction Using Consensus Equilibrium Muhammad Usman Ghani, Student Member, IEEE, W
    SUBMITTED TO IEEE TRANSACTIONS ON COMPUTATIONAL IMAGING 1 Data and Image Prior Integration for Image Reconstruction Using Consensus Equilibrium Muhammad Usman Ghani, Student Member, IEEE, W. Clem Karl, Fellow, IEEE Abstract—Image domain prior models have been shown to and image prior models. Image prior models capture desirable improve the quality of reconstructed images, especially when image features, which enable MBIR methods to produce data are limited. Pre-processing of raw data, through the implicit higher quality image reconstructions [18–20]. A variety of or explicit inclusion of data domain priors have separately also shown utility in improving reconstructions. In this work, a image priors, including Total-Variation (TV) [21], Markov principled approach is presented allowing the unified integration Random Field (MRF) models [22], and deep-learning-based of both data and image domain priors for improved image re- prior models [23] have been explored. Even simple prior construction. The consensus equilibrium framework is extended models such as TV have been shown to greatly improve to integrate physical sensor models, data models, and image image quality, though at the expense of significantly increased models. In order to achieve this integration, the conventional image variables used in consensus equilibrium are augmented computation. An alternative approach has been to focus on with variables representing data domain quantities. The overall transforming the given observed data to better meet the result produces combined estimates of both the data and the assumptions underlying fast conventional analytical recon- reconstructed image that is consistent with the physical models struction methods. In particular, these approaches pre-process and prior models being utilized.
    [Show full text]
  • Automated Analysis of Breast Mri from Traditional Methods Into Deep Learning
    PDF hosted at the Radboud Repository of the Radboud University Nijmegen The following full text is a publisher's version. For additional information about this publication click this link. http://hdl.handle.net/2066/206297 Please be advised that this information was generated on 2021-10-08 and may be subject to change. M. U. DALMIŞ U. M. INVITATION AUTOMATED ANALYSIS to attend the public defense of my PhD thesis AUTOMATED ANALYSIS OF BREAST MRI BREAST OF ANALYSIS AUTOMATED OF BREAST MRI FROM TRADITIONAL METHODS AUTOMATED ANALYSIS INTO DEEP LEARNING OF BREAST MRI FROM TRADITIONAL METHODS INTO DEEP LEARNING Thursday 12 September 2019, at 12.30 in the Aula of the Radboud University, Comeniuslaan 2, 6525 HP Nijmegen. You are cordially invited to the reception after the defense ceremony PARANYMPHS FROM TRADITIONAL METHODS INTO DEEP LEARNING DEEP INTO METHODS TRADITIONAL FROM Jan-Jurre Mordang [email protected] Jules Jacobs [email protected] MEHMET UFUK DALMIŞ [email protected] MEHMET UFUK DALMIŞ Automated Analysis of Breast MRI: from Traditional Methods into Deep Learning Mehmet Ufuk Dalmı¸s This book was typeset by the author using LATEX2". Book by: Mehmet Ufuk Dalmı¸s Cover design by: Wendy Schoneveld & Mehmet Ufuk Dalmı¸s The research described in this thesis was carried out at the Diagnostic Image Analysis Group, Radboud University Medical Center (Nijmegen, the Netherlands). This work was supported by the European Union's Seventh Framework Programme for research, technological development and demonstration (grant agreement no. 601040). Financial support for publication of this thesis was kindly provided by Radboud Uni- versity Medical Center.
    [Show full text]
  • Image Reconstruction Is a New Frontier of Machine Learning — Editorial for the Special Issue “Machine Learning for Image Reconstruction”
    Image Reconstruction Is a New Frontier of Machine Learning — Editorial for the Special Issue “Machine Learning for Image Reconstruction” Ge Wang, Jong Chu Ye, Klaus Mueller, Jeffrey A. Fessler May 2, 2018 1 Introduction Over past several years, machine learning, or more generally artificial intelligence, has generated overwhelming research interest and attracted unprecedented public attention. As tomographic imaging researchers, we share the excitement from our imaging perspective [1], and organized this special issue dedicated to the theme of “Machine Learning for Image Reconstruction”. This special issue is a sister issue of the special issue published in May 2016 of this journal with the theme “Deep Learning in Medical Imaging” [2]. While the previous special issue targeted medical image processing/analysis, this special issue focuses on data-driven tomographic reconstruction. These two special issues are highly complementary, since image reconstruction and image analysis are two of the main pillars for medical imaging. Together we cover the whole workflow of medical imaging: from tomographic raw data/features to reconstructed images and then extracted diagnostic features/readings. In perspective, computer vision and image analysis are great examples of machine learning, es- pecially deep learning. While computer vision and image analysis deal with existing images and produce features of these images (images to features), tomographic reconstruction produces images of internal structures from measurement data which are various features (line integrals, harmonic com- ponents, etc.) of the underlying images (features to images). Recently, machine learning, especially deep learning, techniques are being actively developed worldwide for tomographic reconstruction, as clearly evidenced by the high-quality papers included in this special issue.
    [Show full text]