Author ORCID Identifier
0000-0002-4350-9910
Date of Graduation
8-2023
Document Type
Dissertation (PhD)
Program Affiliation
Medical Physics
Degree Name
Doctor of Philosophy (PhD)
Abstract
Background
Gastrointestinal cancers exhibit a high mortality rate compared to other cancer types. Among these, pancreatic cancer ranks as the fourth leading cause of cancer-related deaths worldwide. The five-year survival rate remains alarmingly low at a mere 9%. Hepatocellular carcinoma (HCC), another aggressive form of cancer, is rapidly becoming the primary cause of cancer-related deaths in the United States. The treatment of both liver cancer and pancreatic cancer heavily relies on a multidisciplinary approach. Innovative treatment strategies involving dose-escalated regimens, such as stereotactic body radiation therapy (SBRT), are emerging as an important pillar of the management of liver and pancreatic cancer. The success of these treatment modalities hinges upon the precise and standardized segmentation of organs-at-risk and target volumes to ensure the optimal quality of treatment plans.
Methods
We first developed an automated organs-at-risk segmentation tool for upper abdominal radiation therapy treatment. A dataset of 70 patients was collected and utilized as the training set and benchmark for our auto-segmentation tool. We employed the adaptive nnU-Net architecture to develop a model ensemble capable of contouring various organs, including the duodenum, small bowel (ileum and jejunum), large bowel, liver, spleen, kidneys, and spinal cord. The performance of the segmentation tool was evaluated on 75 patients using both contrast-enhanced and non-contrast-enhanced CT images, employing a five-point Likert scale assessment by five experts from three different institutions. To capture contours requiring major edits, we developed a distance-based quality assurance (QA) system. This system identified CT scans that were likely to yield suboptimal contours requiring time-consuming major edits. Evaluation of the QA system was conducted on clinical CT scans, with the clinical review score serving as the ground truth. For target volume segmentation, we employed transformer-based architectures, leveraging self-supervised learning and uncertainty estimation techniques to enhance performance and allow for stylistic customization. A total of 3094 unlabeled CT scans from liver cancer patients, along with 5050 publicly available CT scans, were collected for self-supervised pretraining in liver tumor segmentation. The pretrained encoders were then utilized to optimize downstream liver tumor segmentation models, evaluating the impact of self-supervised learning on tumor segmentation performance. For pancreatic tumor segmentation, we developed an ensemble-based approach incorporating multiple segmentation styles. Probability thresholding was employed to generate the final segmentation, enabling customization according to clinicians' preferences.
Results
Our organs-at-risk segmentation tool achieved a clinical acceptance rate of over 90% for all organs except the duodenum, demonstrating its accuracy in delineation. Quantitative results were comparable to state-of-the-art methods, using a small but high-quality dataset. The QA system achieved an AUC of 0.89 for capturing contours requiring major edits on randomly sampled clinical CT scans. In liver tumor segmentation, our study revealed that self-supervised learning demonstrated 4-5% performance improvement when diverse unlabeled data were used for pretraining. This finding highlights the importance of incorporating a wide range of data during the pretraining stage. For pancreatic tumor segmentation, our ensemble-based segmentation method proved highly effective. It provided pixel-by-pixel uncertainty estimates and allowed customization through probability thresholding. Our customized contours surpassed the performance of the state-of-the-art segmentation model, even when utilizing identical training data, pretraining techniques, and hyperparameters.
Conclusion
Our auto-segmentation system for organs-at-risk achieved high clinical acceptance rates in upper-abdominal radiation treatment. The accompanying QA tool effectively captured contours requiring major edits. Leveraging a wide range of unlabeled data in self-supervised learning improved the performance of our transformer-based segmentation system. Additionally, our uncertainty-guided segmentation network allowed customization and identification of low-confidence regions. Our suite of auto-segmentation tools for pancreatic and liver cancer radiation treatment has the potential to streamline clinical workflows while prioritizing patient safety.
Keywords
Radiation therapy, Deep learning, Auto-segmentation, Liver Cancer, Pancreatic Cancer