Evaluation of Patient-Specific Autosegmentation Via Intentional Overfitting: Assessing Lightweight Architectures for Accelerated Online Workflow In Mrgart for Prostate Cancer
Abstract
Purpose
Deep learning-based organ contouring is essential for optimizing the MR-guided adaptive radiotherapy (MRgART) workflow. However, online contouring remains challenging because of interfractional anatomical variations. This study demonstrated the effectiveness of patient-specific models (PSMs) created by intentionally overfitting treatment-planning MR imaging (MRI). Specifically, whether lightweight network architectures could shorten training times, without compromising segmentation accuracy, to enhance the clinical feasibility of online daily adaptation was investigated.
Methods
A 3D-UNet was trained on the LUND-PROBE dataset (425 prostate cancer MRI cases) as a base model (BM). A PSM was developed by fine-tuning the BM using a single patient's planning MRI and ground-truth contours until the model fully converged. To optimize the workflow for clinical time constraints, this study compared it to a lightweight version (LW-UNet) with fewer parameters. Both models were evaluated using daily MR images from five clinical prostate cancer cases treated with a 1.5T MR-Linac (Unity, Elekta) following MRI simulation. Accuracy was measured using the Dice similarity coefficient (DSC) for the rectum, bladder, and clinical target volume (CTV).
Results
For the standard 3D-UNet, the PSM generally improved or maintained high mean DSCs compared to the BM: rectum (0.77 to 0.86), bladder (0.90 to 0.90), and CTV (0.64 to 0.90). The LW-UNet PSM yielded comparable high-accuracy results, achieving mean DSCs of 0.86 for the rectum, 0.87 for the bladder, and 0.91 for the CTV. Notably, the LW-UNet substantially reduced the training duration required for overfitting compared with the standard 3D-UNet (24.7 min vs. 55.8 min), facilitating a more rapid model generation process.
Conclusion
Intentional overfitting of pretreatment planning data effectively created high-performance PSMs. These findings suggest that lightweight architectures provide a critical balance between computational efficiency and segmentation precision. This approach meaningful shortens the patient-specific training phase, making real-time personalized auto-segmentation a practical solution for online MRgART workflows.