1072 - Patient-Specific Priors for CTV Auto-Contouring in Online Adaptive Radiotherapy for Cervical Cancer

Purpose/Objective(s): Cervical cancer online adaptive radiotherapy (oART) is often limited by the time required to generate clinically acceptable target contours. We hypothesized that anchoring a population segmentation model with patient-specific CT and contour priors would improve accuracy and reduce manual editing, including under anatomical distribution shift and limited training data.

Materials/Methods: A 3D U-Net general model (GM) was trained on pelvic simulation CTs (pelvic IMRT cohort n=96; 76 train, 20 validate) and evaluated on the 20th fraction CT. A personalized model (PM) was derived per patient by one-shot or few-shot fine-tuning using the planning CT and physician contours; for oART, PMs additionally leveraged accumulating prior-fraction FBCT contours. Comparators were deformable image registration (DIR) and deformable registration-guided deep learning (Def-RgDL). Generalizability was tested in extended-field IMRT including para-aortic nodes (n=47; 39 train, 8 test). Clinical impact was assessed in fan-beam computed tomography (FBCT)-guided oART (n=20; 15 train, 5 evaluate; 50 fractions) across three PM update schedules. An experienced radiation oncologist edited auto-contours blinded to method; editing time from contour availability to approval was recorded. Endpoints were DSC, HD95, ASD, and editing time using paired statistical tests.

Results: In pelvic validation, PM consistently outperformed DIR, GM, and Def-RgDL in geometric accuracy, achieving a mean DSC of 0.925 with lower surface distance errors; these improvements were statistically significant. Under extended-field distribution shift, PM maintained strong performance (DSC = 0.934), whereas direct transfer of the pelvic GM showed substantial degradation. In data-efficiency analyses, personalization preserved high accuracy with fine-tuning on as few as 7 training patients, exceeding both population and registration-guided baselines. In FBCT-guided oART, a weekly updated PM using all prior fractions delivered stable multi-target performance (DSC 0.90 to 0.94 across CTV subregions) and reduced physician editing time by approximately 60% compared with GM.

Conclusion: Patient prior anchored personalization provides a data-efficient route to deploy accurate cervical cancer CTV auto-contouring in FBCT-guided oART, improves robustness to distribution shift, and meaningfully reduces physician editing burden. This supports broader adoption of adaptive workflows.