Test‐time adaptation‐driven patient‐specific auto‐segmentation for CBCT‐guided online adaptive radiotherapy in cervical cancer

BACKGROUND: Online adaptive radiotherapy using cone beam computed tomography (CBCT) of the pelvis can account for anatomical variations during cervical cancer treatment. However, the limitations of CBCT and existing auto-segmentation methods, such as deformable image registration (DIR), hinder contouring and dosimetric accuracy. PURPOSE: We propose a test-time adaptation (TTA)-driven patient-specific segmentation framework to enhance the accuracy and workflow efficiency of CBCT-guided online adaptive radiotherapy in cervical cancer. METHODS: Our retrospective analysis included 20 patients with cervical/endometrial cancer (530 CBCT scans) treated on the Varian Ethos platform. A patient-specific UNet variant pre-trained on the CTPelvic1K dataset using Bootstrap Your Own Latent (BYOL) contrastive learning was fine-tuned on individual reference CT scans. During daily adaptation, the model incorporated TTA with segmentation (cross entropy + Dice loss) and consistency objectives, leveraging prior CBCT/CT images to constrain anatomical variations. Performance was evaluated using the Dice similarity coefficient (DSC), 95th percentile Hausdorff Distance (HD95), and contouring time reduction compared to DIR-based methods. RESULTS: The proposed method significantly outperformed DIR, achieving higher DSC (e.g., postoperative clinical target volume (CTV) of the vaginal cuff and upper vaginal region: 0.89 vs. 0.73; definitive CTV of the pelvic lymphatic drainage area: 0.83 vs. 0.71) and lower HD95 (e.g., rectum: 24.9 vs. 30.7 mm) values. TTA further improved the accuracy, particularly for the bowel (DSC: 0.81 vs. 0.76) and target volumes. Clinically, the proposed method reduced the contouring time by 250 s (postoperative) and 230 s (definitive) per fraction (p < 0.01), with reduced time variability. Structures with clear boundaries required one or two fractions for training, whereas motion-prone organs required four or five fractions. CONCLUSIONS: This study presents the first retrospective demonstration of TTA-driven patient-specific segmentation for pelvic CBCT-guided online adaptive radiotherapy. The proposed framework enhances segmentation accuracy, workflow efficiency, and temporal consistency by addressing the inherent challenges and data scarcity of CBCT through adaptive learning.