2026-05-20T23:29:42Zhttps://keep.lib.asu.edu/oai/request

oai:keep.lib.asu.edu:node-1992032024-12-23T18:01:48Zoai_pmh:repo_items

199203 https://hdl.handle.net/2286/R.2.N.199203 http://rightsstatements.org/vocab/InC/1.0/ All Rights Reserved 2024 2026-12-01T12:21:06 111 pages Masters Thesis Academic theses Text eng Saravanan, Madhumitha Liang, Jianming Gopalan, Nakul Davulcu, Hasan Arizona State University Partial requirement for: M.S., Arizona State University, 2024 Field of study: Computer Science Deep learning has transformed healthcare by enhancing diagnostic, surgical, and therapeutic capabilities through automated organ and tumor segmentation. Nevertheless, annotating datasets for training deep learning models remains labor-intensive, often resulting in small, inconsistent datasets. This constraint has led to the development of numerous single-task models tailored to specific segmentation tasks, limiting adaptability to new labels. The absence of a contemporary benchmark comparing recent advancements with long-established models further complicates model selection. In this study, the state-of-the-art single-task models were compared, identifying MedNeXt and U-Mamba_Bot as top performers in 3D segmentation, especially with limited data. Moreover, the heterogeneous nature of medical imaging datasets, with varied annotation distributions, presents challenges in developing robust, extensible models. Existing techniques that aim to integrate knowledge from different tasks often struggle with catastrophic forgetting, risks to data integrity from label adjustments, and inability to scale to new tasks. To combat these difficulties, the Cyclic, Lock, and Release pre-training strategies were employed, leveraging a shared encoder to integrate knowledge across tasks while incorporating task-specific components to address unique attributes of each task. The resulting model, which was pre-trained on 16 public datasets having 3,000 CT scans with annotations for 25 organs and 6 tumors, outperformed both the single-task Swin UNETR and the multi-task CLIP-driven Universal model. It exhibited superior generalization and data efficiency when fine-tuned on the TotalSegmentator dataset. This exceptional performance is attributed to its ability to effectively learn from expert annotations without requiring alterations, while also preventing bias towards a single task and mitigating catastrophic forgetting. This positions the model employing the Cyclic, Lock, and Release pre-training strategies as a strong competitor to current state-of-the-art solutions. Computer Science 3D Segmentation Computer Aided Diagnosis Computer vision Deep learning Multi-modal Learning Multi-Task Learning Benchmarking and Boosting of 3D Segmentation Models