J. Hwang1, H. Eum2, J. S. Kim3, E. Lee4, B. H. Kang5, and Y. Park5; 1Department of Nuclear and Quantum Engineering, Korea Advanced Institute of Science and Technology(KAIST), Daejeon, Korea, Republic of (South), 2Medical Physics and Biomedical Engineering Laboratory (MPBEL), Yonsei University College of Medicine, Seoul, Korea, Republic of (South), 3Department of Radiation Oncology, Yonsei Cancer Center, Yonsei University College of Medicine, Seoul, Korea, Republic of (South), 4Ewha Womans University Medical Center, Seoul, Korea, Republic Of, 5Ewha, Seoul, Korea, Republic of (South)
Purpose/Objective(s): Stereotactic Body Radiation Therapy (SBRT) presents challenges in dose planning due to daily variations in organ and tumor structures. Online adaptive radiation therapy (ART) has been introduced to address these challenges, requiring rapid and accurate auto-segmentation and planning. Precise delineation of the Gross Tumor Volume (GTV) is crucial for treatment outcomes, often necessitating manual adjustments. This study employs the Intentional Deep Overfit Learning (IDOL) framework to enhance patient-specific segmentation accuracy in prostate cancer patients undergoing SBRT.Materials/
Methods: A retrospective analysis of 48 Cone Beam Computed Tomography (CBCT) scans from 10 prostate cancer patients was conducted. Each patient contributed 4 to 5 fractions of CBCT scans. Utilizing the Swin UNETR deep learning model, an innovative methodology was developed. Eight patients underwent pre-training with 38 CBCT scans, followed by intentional overfitting techniques within the IDOL framework applied to the remaining 2 patients to generate highly patient-specific models. Results: Evaluation using the Dice Similarity Coefficient (DSC) across the initial five treatment fractions demonstrated significant improvements in GTV segmentation compared to a pre-trained network. Application of the IDOL framework during the 1st fraction increased the average DSC from 0.9246 to 0.9598 in GTV segmentation, indicating precision gains. PTV segmentation also improved, with the DSC rising from 0.9428 to 0.9501 after IDOL application during the 1st fraction, outperforming the pre-trained network. The IDOL framework exhibited efficiency with a brief training time averaging approximately 5 minutes per patient-specific model, highlighting its practicality in clinical settings. Conclusion: The IDOL framework demonstrated notable improvements in GTV segmentation and modest enhancements in PTV segmentation, underscoring its effectiveness in refining accuracy in prostate SBRT planning. This research contributes to the evidence supporting the versatile applicability of the IDOL framework, showcasing its potential to enhance patient-specific segmentation precision and ultimately improve treatment outcomes in adaptive SBRT for prostate cancer patients.
