Label-efficient machine learning for medical image analysis
Abstract/Contents
- Abstract
- Medical imaging is an essential tool in healthcare, and radiologists are highly trained to detect and characterize disease in medical images. However, relying solely on human analysis has limitations: it can be time consuming, variable, and difficult to scale. Automating portions of the medical image analysis pipeline can overcome these limitations to support and expand the capabilities of clinicians and radiologists. In this dissertation, we focus on the potentially transformative role deep learning will play in automated medical image analysis. We pose segmentation as a key tool for deep learning-based image analysis, and we show how segmentation neural networks can achieve high performance on many medical image analysis tasks without large, manually annotated training datasets. We begin by describing two methods for training medical image segmentation neural networks with limited labeled data. In our first method, we adapt weak supervision to segmentation. In our second method, we fuse data augmentation, consistency regularization, and pseudo labeling in a unified semi-supervision pipeline. These methods fold multiple approaches to limited-label training into the same framework, leveraging the strengths of each to achieve high performance while keeping labeling burden low. Next, we evaluate networks trained with limited labeled data on clinically motivated metrics over multi-institution, multi-scanner, multi-disease datasets. We find that our semi-supervised networks achieve improved performance compared to fully supervised networks (trained with over 100x more labeled data) on certain generalization tasks, achieving stronger concordance with a human annotator. However, we uncover data subsets on which the label-efficient methods underperform. We propose an active learning extension to our semi-supervised pipeline to address these error modes, improving semi-supervised performance on a difficult data slice by 18.5%. Through this evaluation, we develop an understanding of how networks trained with limited labeled data perform on clinical tasks, how they compare to networks trained with abundant labeled data, and how to mitigate error modes. Finally, we apply label-efficient segmentation models to a broader set of medical image analysis tasks. Specifically, we demonstrate how and why segmentation can benefit medical image classification. We first analyze why segmentation versus classification models may achieve different performances on the same dataset and task. We then implement methods for using segmentation models to classify medical images, which we call segmentation-for-classification, and compare these methods against traditional classification on three retrospective datasets. Finally, we use our analysis and experiments to summarize the benefits of using segmentation-for-classification compared to standard classification, including: improved sample efficiency, enabling improved performance with fewer labeled images (up to an order of magnitude fewer), on low-prevalence classes, and on certain rare subgroups (up to 161.1% improved recall); improved robustness to spurious correlations (up to 44.8% improved robust AUROC); and improved model interpretability, evaluation, and error analysis. These results show that leveraging segmentation models can lead to higher-quality medical image classifiers in common settings. In summary, this dissertation focuses on segmentation as a key tool for supporting automated medical image analysis, and we show how to train segmentation networks to achieve high performance on many image analysis tasks without large labeling burdens.
Description
| Type of resource | text |
|---|---|
| Form | electronic resource; remote; computer; online resource |
| Extent | 1 online resource. |
| Place | California |
| Place | [Stanford, California] |
| Publisher | [Stanford University] |
| Copyright date | 2023; ©2023 |
| Publication date | 2023; 2023 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Hooper, Sarah McIlwaine |
|---|---|
| Degree supervisor | Langlotz, Curtis P |
| Degree supervisor | Ré, Christopher |
| Thesis advisor | Langlotz, Curtis P |
| Thesis advisor | Ré, Christopher |
| Thesis advisor | Nishimura, Dwight George |
| Thesis advisor | Olukotun, Oyekunle Ayinde |
| Degree committee member | Nishimura, Dwight George |
| Degree committee member | Olukotun, Oyekunle Ayinde |
| Associated with | Stanford University, School of Engineering |
| Associated with | Stanford University, Department of Electrical Engineering |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Sarah Hooper. |
|---|---|
| Note | Submitted to the Department of Electrical Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2023. |
| Location | https://purl.stanford.edu/mx843ys4683 |
Access conditions
- Copyright
- © 2023 by Sarah McIlwaine Hooper
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...