3018 - Clinical Impact of Multi-Task Learning and On-Site Fine-Tuning for Auto-Segmentation In Left-Sided Breast Cancer Radiotherapy
Presenter(s)
E. J. Heo1, S. H. Cho2, D. Lee3, K. H. Chang4, J. B. Shim5, N. K. Lee6, C. Y. Kim6, and S. Lee6; 1Department of Medical Physics, Graduate School of Korea University, Sejong, Korea, Republic of (South), 2Korea University Hospital Guro Hospital, Dept Radiation Oncology, Seoul, Korea, Republic of (South), 3Department of Sales and CS, OncoSoft, Seoul, Korea, Republic of (South), 4Department of Radiologic Science, Far East University, Chungcheongbuk-do, Korea, Republic of (South), 5Department of Medical Physics, Kyonggi University, Suwon-si, Korea, Republic of (South), 6Department of Radiation Oncology, College of Medicine, Korea University, Seoul, Korea, Republic of (South)
Purpose/Objective(s): Auto-segmentation models often require substantial manual editing in clinical practice due to inter-institutional variations in imaging protocols and contouring practices. This study directly compared a multi-task learning (MTL)-based model and an on-site fine-tuned model against a commercial pre-built model to evaluate their impact on geometric accuracy, boundary robustness, and physician satisfaction in postoperative left-sided breast cancer radiotherapy.
Materials/Methods: A retrospective analysis of 119 postoperative left-sided breast cancer patients treated with simultaneous integrated boost (SIB) radiotherapy was conducted. An MTL model was developed to simultaneously perform auto-segmentation and dose prediction tasks, leveraging common information between these interdependent tasks. Additionally, commercial pre-built model was fine-tuned using institutional data to generate an on-site trained model. All three models (pre-built, on-site trained, and MTL) were trained and evaluated using an identical dataset configuration to ensure fair comparison. Geometric accuracy was assessed for the planning target volume (PTV), boost volume, heart, left lung, and right lung using the Dice similarity coefficient (DSC) and the 95th percentile Hausdorff distance (HD95). Physician satisfaction was evaluated through structured, blinded assessment of heart auto-segmentation focusing on clinical critical regions including the borders, chambers, great vessels, and coronary arteries.
Results: Compared with the pre-built model, both the on-site trained and MTL models significantly improved PTV auto-segmentation accuracy. The on-site trained model increased the DSC for PTV from 0.664 to 0.903, while the MTL model achieved a DSC of 0.886 (both p < 0.001). HD95 was significantly reduced from 38.5 mm in the pre-built model to 9.3 mm (on-site trained model) and 5.5 mm (MTL model) (both p < 0.001), indicating substantially improved boundary definition . In the physician-blinded evaluation, the on-site trained model did not demonstrate a significant difference in acceptance ratio compared with the pre-built model. In contrast, the MTL model significantly improved physician acceptance in the cranial border and great vessels regions. The acceptance ratio for great vessels increased substantially from 32.9% to 76.1%, and cranial border acceptance ratio was similarly enhanced. Heart auto-segmentation acceptance demonstrated the most marked improvement with the MTL model.
Conclusion: The on-site trained model effectively optimized volume overlap metrics; however, the MTL model demonstrated superior boundary robustness and substantially reduced physician editing burden, particularly for cardiac and vascular structures. These findings support the clinical applicability of the MTL approach, suggesting its potential to streamline auto-segmentation workflows while maintaining clinical quality.