Date of Award
Doctor of Philosophy (PhD)
The applications of deep learning (DL) in medical imaging have been extensively developed and investigated to tackle a wide variety of problems, with many of them able to achieve state-of-the-art performances in different tasks. In the field of radiation therapy (RT), the workflow is a complex process consisting of multiple time and effort-consuming steps such as organ-at-risk (OAR) contouring that could impact treatment quality and patient outcomes. Due to the rapid growth in patient data and the advanced DL techniques, increased attention has been paid to the clinical application of machine learning (ML) and DL1,2 and commercial software has been integrated into clinical workflows in some applications. However, in a clinical setting, image appearance can change due to various factors that yielding data domain shift, and how the trained DL models respond to the shifted clinic data remains unknown without ground truth labels. This work seeks to fill the gaps between DL auto-segmentation algorithm development and clinical deployment, which can be summarized as follows. 1) We lack a comprehensive understanding of the robustness of DL auto-segmentation algorithms against domain changes in a clinical environment. 2) We are in lack of a method to quantify the algorithm robustness given test data from various domains. 3) Currently, clinically useable QA tools and benchmark datasets to monitor the clinical performance of AI algorithms are missing. Firstly, a DL based cardiac substructure segmentation algorithm was developed, which was in turn used as a well-understood example problem and baseline to investigate a framework of robustness and quality assurance (QA). Then a QA framework for monitoring the performance of the DL model was proposed, including three components: 1) a comprehensive benchmark dataset to evaluate the model robustness and limitation. 2) Autoencoder based techniques and hand-engineered features for image domain shift quantification and segmentation shape assessment. 3) a regression model that takes in the extracted features and makes clinically acceptable predictions of segmentation performances. Finally, the proposed QA framework was applied to commercial auto segmentation software and demonstrated relatively good performance.
Yao Hao, Abhinav Jha, Yuan-Chuan Tai, Tiezhi Zhang,
Available for download on Thursday, October 26, 2023