Author: Seungryong Cho, Donghyeok Choi, Joonil Hwang, Byung-Hee Kang, Jin Sung Kim, Eungman Lee, Younghee Park 👨🔬
Affiliation: Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, KAIST, Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Ewha Womans University of Medicine 🌍
Purpose: Radiation therapy (RT) is critical for cancer treatment, but changes in tumor size and shape during therapy challenge precise dose delivery. Adaptive radiation therapy (ART) addresses these variations by requiring accurate segmentation of Gross Tumor Volume (GTV) and Planning Target Volume (PTV). This study introduces the Intentional Deep Overfit Learning (IDOL) framework, utilizing IDOLadaptive and IDOLsequence strategies to enhance segmentation accuracy for prostate cancer ART.
Methods: The IDOL framework employs the Swin UNETR model, integrating transformer-based and convolutional neural network architectures for volumetric segmentation. In the IDOLadaptive strategy, the model is iteratively fine-tuned using data from the first treatment fraction and sequentially incorporating additional fractions, testing on the fifth fraction. The IDOLsequence strategy fine-tunes the model incrementally with temporally adjacent fractions, adapting to anatomical changes. Both strategies were evaluated against pre-trained models and traditional deformable image registration techniques using metrics like Dice Similarity Coefficient (DSC), Hausdorff Distance(HD), and Mean Surface Distance(MSD).
Results: For GTV segmentation in the fifth fraction, IDOLadaptive achieved a DSC of 0.9784, outperforming IDOLsequence (0.9744) and the pre-trained model (0.9324). For PTV segmentation, IDOLadaptive also excelled with a DSC of 0.9501. Significant improvements were noted in the 95th percentile HD and MSD, confirming the precision of the IDOL framework.
Conclusion: The IDOL framework dynamically adapts to sequential CBCT data, refining segmentation accuracy by leveraging patient-specific data. IDOLadaptive demonstrated superior performance through progressive fine-tuning, while IDOLsequence provided robust segmentation by focusing on temporally adjacent data. This approach enhances ART by addressing inter-fractional anatomical changes, offering a transformative solution for personalized prostate cancer treatment planning.