Automated longitudinal assessment of CT cancer lesions
Abstract/Contents
- Abstract
- Objective assessment of cancer imaging studies remains a gap in patient care. Although lesion assessment is a critical part of assessing cancer treatment response, in current clinical practice it is mostly done qualitatively through free-text reports. Quantitative lesion assessment is usually only done in clinical trials, representing a small fraction of all cancer cases. To address this need, this thesis presents an automated pipeline for longitudinal assessment of cancer lesions. Given an annotated baseline exam, the system automatically detects, segments and measures the corresponding target lesions in one or more follow-up exams. It also learns from prior time points by training on partial lesion annotations made by physicians. The core components of this system are neural networks for organ and lesion segmentation, image registration from 3D SIFT-like keypoints, and a method for combining all of these into an end-to-end system that learns from prior time points. This thesis presents several contributions to each of these technical areas. For organ segmentation, we present a new dataset labeling six organs in a variety of CT scans. Despite the relative ease of locating organs in the human body, automated organ segmentation has been hindered by the scarcity of labeled training data. Due to the tedium of labeling organ boundaries, most datasets are limited to either a small number of cases or a single organ. Furthermore, many are restricted to specific imaging conditions unrepresentative of clinical practice. To address this need, we developed a diverse dataset of 140 CT scans containing six organ classes: liver, lungs, bladder, kidney, bones and brain. For the lungs and bones, we expedited annotation using unsupervised morphological segmentation algorithms, which were accelerated by 3D Fourier transforms. We trained a deep neural network which requires only 4.3 s to simultaneously segment all the organs in a CT exam. We also show how to efficiently augment the data to improve model generalization, providing a GPU library for doing so. The dataset and code are available through The Cancer Imaging Archive. For lesion segmentation, we show how the same neural network architecture can be generalized to segment lesions using public datasets with CT scans of two different organs: the liver and the lungs. For our ultimate goal of longitudinal segmentation, we also show how to train these models on partially labeled CT scans from prior timepoints in the clinical record. For image registration, we present a method based on 3D scale- and rotation-invariant keypoints. The method extends the Scale Invariant Feature Transform (SIFT) to arbitrary dimensions by making key modifications to orientation assignment and gradient histograms. Rotation invariance is proven mathematically. Additional modifications are made to extrema detection and keypoint matching based on the demands of image registration. Our experiments suggest that the choice of neighborhood in discrete extrema detection has a strong impact on image registration accuracy. In head MR images, the brain is registered to a labeled atlas with an average Dice coefficient of 92%, outperforming registration from mutual information as well as an existing 3D SIFT implementation. In abdominal CT images, the spine is registered with an average error of 4.82 mm. Furthermore, keypoints are matched with high precision in simulated head MR images exhibiting lesions from multiple sclerosis. These results were achieved using only affine transforms, and with no change in parameters across a wide variety of medical images. This work is freely available as a cross-platform software library. Finally, we combine all these technologies into and end-to-end system for tracking cancer lesions over time in CT images. We tested this system on a retrospective cohort of clinical trial lesion annotations which was automatically curated from our institutional record. Our retrospective study involved 407 patients (51% male and 49% female) with exam dates ranging from 05/02/2006 -- 09/30/2020. Given an annotated baseline exam, we used an automatic pipeline to detect, segment and measure lesions in subsequent exams. We compared our lesion measurements to those made in clinical trials, reporting statistics on lesion detection and measurement errors, as well as the improvement from training on prior exams. When tracking a labeled a baseline lesion in a follow-up exam, our system detected the correct follow-up lesion for 55.5% of liver and 49.6% of lung annotations. In other cases, the system either identified a different lesion (37.9% liver, 49.6% lung), or found no corresponding lesions in the follow-up (6.6% liver, 0.8% lung). The median measurement error for correctly-detected lesions was 3.47 mm (95% CI: [2.22, 5.67]) for liver and 1.36 mm (95% CI: [0.97, 2.03]) for lung. This resulted in 83.9% of exams receiving a correct tumor response grade for liver and 90.4% for lung, using the Response Evaluation Criteria in Solid Tumors (RECIST) guidelines. Training on prior exams improved both detection (p = .000024) and segmentation (p = .018) in the liver, while degrading both detection (p = .00019) and segmentation (p = .018) in the lung. Overall, our system automates detection and measurement of roughly half of all target lesions across multiple organs. As the main contribution is heavily dependent on deep learning, this thesis also contains a pair of appendices presenting original results on neural network theory. The first discusses convexity properties of the training loss function, and the second discusses the effect of model architecture on neuron death, studied through random initialization of model parameters. The appendix proves some local convexity properties of the training loss and inference functions, as well as upper and lower bounds on the probability that a ReLU network is initialized to a trainable point, as a function of model hyperparameters. It also contains some proposals for initialization and local optimization strategies based on these results.
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 | 2022; ©2022 |
| Publication date | 2022; 2022 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Author | Rister, Blaine Burton |
|---|---|
| Degree supervisor | Rubin, Daniel (Daniel L.) |
| Thesis advisor | Rubin, Daniel (Daniel L.) |
| Thesis advisor | Duchi, John |
| Thesis advisor | Horowitz, Mark (Mark Alan) |
| Degree committee member | Duchi, John |
| Degree committee member | Horowitz, Mark (Mark Alan) |
| Associated with | Stanford University, Department of Electrical Engineering |
Subjects
| Genre | Theses |
|---|---|
| Genre | Text |
Bibliographic information
| Statement of responsibility | Blaine Burton Rister. |
|---|---|
| Note | Submitted to the Department of Electrical Engineering. |
| Thesis | Thesis Ph.D. Stanford University 2022. |
| Location | https://purl.stanford.edu/yx087mh5697 |
Access conditions
- Copyright
- © 2022 by Blaine Burton Rister
- License
- This work is licensed under a Creative Commons Attribution 3.0 Unported license (CC BY).
Also listed in
Loading usage metrics...