Dissertation
Deep learning and explainable AI in medical imaging
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Spring 2023
DOI: 10.25820/etd.007288
Abstract
Medical imaging is a critical part of modern healthcare, and the continued advancement of imaging techniques have enabled the analysis of an every growing number of objects and features within the human body that we were previously unable to observe. However, the significant variability present in imaging protocols and the features they capture make the growing number of images generated in the process of daily patient care a mounting burden on the limited number of radiological experts available to process them. Deep learning techniques have emerged as powerful tools for the efficient and accurate automation of medical image segmentation, but the complexity of the models required to handle this difficult domain severely limits their transparency and therefor trustworthiness to the end users. To address these concerns and build the trust necessary for clinical implementation, numerous methods have been proposed to derived explanations for the decisions of deep learning models. Unfortunately, these efforts have almost exclusively targeted models developed for classification tasks and their underlying assumptions prevent them from applying to segmentation.
In the first part of this thesis, we develop high-quality segmentation approaches for a diverse range of medical imaging targets. Specifically, we train deep learning models for the segmentation of thoracic organs at risk during radiotherapy, pulmonary tumors associated with non-small cell lung cancer, COVID-19 related lung lesions, and the prostate and surrounding organs. For each of these tasks, we utilize state-of-the-art hybrid transformer models paired with powerful self-configuring preprocessing schemes to achieve highly accurate and consistent segmentations that closely align with the independent standards. The second part of this thesis focuses on producing and understanding visual explanations for these kinds of models. We develop and validate our Kernel-Weighted Contribution approach to explaining the complex features driving the decisions made by deep learning segmentation models. We demonstrate that our approach offers accurate and comprehensive explanations that enable greater understanding of these tools.
Building on the successes of the first two parts, the third part of our thesis demonstrates the utility of these approaches in the context of medical imaging analysis. We show that our segmentation models not only enable current forms of medical analyses, such as the extraction of radiomics features that can inform patient care and disease prognosis, but also enable novel applications, such as evaluating contour consistency across time points in radiotherapy planning. We also demonstrate the critical role of our Kernel-Weighted Contribution approach in validating our segmentation models by detecting and evaluating the hidden biases that would otherwise have gone undetected by conventional analysis, ultimately allowing us to gain insight into how our models would behave when they are applied beyond the research lab in which they were developed.
By developing high-quality segmentation approaches to tackle complex medical imaging tasks and proposing a novel method for the visual explanation of deep learning segmentation models, we have demonstrated the potential of deep learning to enable accurate and efficient medical image analyses while also increasing our understanding of the processes used to accomplish those analyses. This kind of understanding represents an important step towards opening the ``black box'' that is deep learning and building the trust necessary for these powerful models to be integrated into clinical practice and provide direct benefits to patient care and outcomes.
Details
- Title: Subtitle
- Deep learning and explainable AI in medical imaging
- Creators
- Sean Mullan
- Contributors
- Milan Sonka (Advisor)Joseph M Reinhardt (Committee Member)Eric A Hoffman (Committee Member)Sajan Lingala (Committee Member)Daniel Hyer (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Biomedical Engineering
- Date degree season
- Spring 2023
- DOI
- 10.25820/etd.007288
- Publisher
- University of Iowa
- Number of pages
- xix, 154 pages
- Copyright
- Copyright 2023 Sean Mullan
- Language
- English
- Date submitted
- 04/24/2023
- Date approved
- 05/12/2023
- Description illustrations
- color illustrations
- Description bibliographic
- Includes bibliographical references (pages 131-148).
- Public Abstract (ETD)
Medical imaging is critical part in modern healthcare, but the increasing number of images generated by routine patient care can quickly overwhelm the limited number of radiologists available for complex tasks. Deep learning is a powerful tool that can accurately and efficiently analyze these images, but the complexity of these models makes it impossible for doctors and patients to understand how decisions are made. Previous work has investigated explaining deep learning classification, providing a single decision for an image, but these approaches don’t apply to deep learning segmentation, classifying each pixel in an input image to separate objects from background.
In the first part of this thesis, we use state-of-the-art approaches to develop high-quality models that accurately segment a wide range of complex targets. Next, we propose a novel method for visual explanation, Kernel-Weighted Contribution, which provides comprehensive explanations enabling understanding of the features that drive these models. Finally, we show that our models can be used for both current and novel medical analyses, and we use our explanation approach to validate our segmentation models by understanding the features that they have learned and how they use them. Our work shows the potential of deep learning to enable accurate medical image analyses and allows improved understanding of these complex models. This kind of understanding represents an important step towards opening the “black box” of deep learning and building the trust necessary for these powerful models to be integrated into clinical practice and provide direct benefits to patient care and outcomes.
- Academic Unit
- Roy J. Carver Department of Biomedical Engineering
- Record Identifier
- 9984424790702771
Metrics
9 File views/ downloads
33 Record Views