Dissertation
Enabling quality assessment for automated medical image segmentation by deep learning, with applications
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Summer 2024
DOI: 10.25820/etd.007699
Abstract
Image segmentation, which divides images into multiple segments at the pixel/voxel level, is essential for a wide range of computer vision applications. In the field of medical imaging, the quantitative analysis of segmented objects using diverse imaging modalities such as magnetic resonance imaging (MRI), computed tomography (CT), optical coherence tomography (OCT), positron emission tomography (PET), Ultrasound, etc., is extensively employed for monitoring disease progressions. This involves characterizing, tracking, and predicting the development of diseases, as well as assessing the impact of pharmacotherapies on treatment responsiveness. Maintaining high-quality segmentation is crucial for ensuring that image-derived quantitative metrics support clinical decision-making in both diagnosis and treatment. Recent advancements in automated image segmentation based on deep learning have introduced numerous efficient methods, achieving accuracies comparable to those of expert-level professionals in the field. Nevertheless, conventional deep learning approaches may encounter challenges in segmenting 3D medical images due to complexities in tissue structures, variability in image acquisition quality, disease-related irregularities, and limitations in the availability of sufficiently large training datasets. There is a compelling need for robust and accurate segmentation quality assessment tools in clinical settings. This research aims to explore the quality assessment methods of automated medical image segmentation, as well as explores robust, accurate and efficient medical image segmentation techniques that can be used in multi-imaging modalities. This study proposes three medical image segmentation quality assessment models from different perspectives: (1) Deep learning-based error detection model on the segmented surface of the target objects with area optimization, (2) Deep learning-based generative-regression model for patch-wise medical image segmentation quality assessment and (3) Deep learning-based verification model to detect adversarial attack for robust quality assessment of the medical image segmentation. This study also proposes two diffusion-based segmentation models: (1) Score-based surface segmentation model by adapting graph re-parameterization and cold-diffusion and (2) Denoising diffusion probabilistic model in latent-space for medical image segmentation. Furthermore, this study investigates a rare cardiovascular disease: Takotsubo Syndrome to develop deep learning based accurate diagnostic models and explores robust pathophysiological feature extraction for differential diagnosis in clinical and urgent care settings. The effectiveness of the models was validated by experimenting data from various imaging modalities, such as, MR images of knee bones and cartilage, calf-muscle compartments, CT images of lung tumor, OCT images of retina, echocardiography (echo) videos of heart etc.
Details
- Title: Subtitle
- Enabling quality assessment for automated medical image segmentation by deep learning, with applications
- Creators
- Fahim Ahmed Zaman
- Contributors
- Xiaodong Wu (Advisor)Milan Sonka (Committee Member)Kan Liu (Committee Member)Mona Garvin (Committee Member)Tyler Bell (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Summer 2024
- DOI
- 10.25820/etd.007699
- Publisher
- University of Iowa
- Number of pages
- xix, 119 pages
- Copyright
- Copyright 2024 Fahim Ahmed Zaman
- Grant note
- Finally, thanks to the funding institutions. This research was supported, in part, by the NIH Grant R01-EB004640, R01-NS094387, R01-AG067078, R56-EB004640, R01-HL171624 and R01-EB019961. The Takotsubo project was also funded by the Obermann Center for Advanced Studies Interdisciplinary Research Grant and Institute for Clinical and Translational Science Grant. (iii)
- Comment
- This thesis has been optimized for improved web viewing. If you require the original version, contact the University Archives at the University of Iowa: https://www.lib.uiowa.edu/sc/contact/
- Language
- English
- Date submitted
- 07/20/2024
- Description illustrations
- Illustrations, tables, graphs, charts
- Description bibliographic
- Includes bibliographical references (pages 103-119).
- Public Abstract (ETD)
Image segmentation is a fundamental task in computer vision, involving the division of images into distinct segments at the pixel/voxel level. In the realm of medical imaging, this technique is widely employed to perform quantitative analyses on segmented objects using various modalities such as magnetic resonance imaging (MRI), computed tomography (CT), optical coherence tomography (OCT), positron emission tomography (PET), Ultrasound, among others. Its applications extend to monitoring disease progression, predicting developments, and evaluating treatment effectiveness. Accurate segmentation is essential to ensure that quantitative metrics derived from images support clinical decision-making in both diagnosis and treatment. Recent advancements in deep learning have introduced highly efficient segmentation methods, achieving accuracies comparable to those of experts in the field. However, segmenting 3D medical images remains challenging due to complex tissue structures, variability in image quality during acquisition, disease-related variations, and limitations in availability of large-scale training datasets. This study addresses the critical need for robust quality assessment tools in clinical settings, focusing specifically on automated medical image segmentation. The study proposes three innovative quality assessment models from different perspectives. Additionally, two diffusion-based segmentation models are proposed for accurate medical image segmentation. Furthermore, the research explores Takotsubo Syndrome, a rare cardiovascular disease, by developing accurate diagnostic models using deep learning techniques. It also investigates robust extraction of pathophysiological features to facilitate differential diagnosis in clinical and urgent care settings. Validation experiments are conducted using diverse imaging modalities, including MR images of knee bones and cartilage, calf-muscle compartments, CT images for lung tumors, OCT images of retina, and echocardiography videos of heart.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9984698053402771
Metrics
1 File views/ downloads
9 Record Views