3178 - Feasibility of Online Adaptive Radiotherapy for Bladder Cancer Using Deep Learning Auto-Contouring and Auto-Planning on CBCT

Purpose/Objective(s):

Bladder shape and size at treatment often differ substantially from simulation CT, typically necessitating large PTV margins (e.g., 1 cm). We hypothesized that online adaptive radiotherapy (ART) using deep learning (DL) to automatically contour and plan on cone-beam CT (CBCT) would be feasible and would improve target coverage and reduce organ-at-risk (OAR) doses compared with conventional non-adaptive plans.

Materials/Methods:

An IRB-approved retrospective study included 16 bladder cancer cases (IRB2025-00013); exclusions included body habitus beyond CBCT field of view, Foley catheter or balloon, and significant calcification-related artifact. Bladder, bowel, and rectum were contoured on the CBCT images acquired at the first and last fractions over the entire course of treatment. A UNet model was trained on first-CBCT contours and applied to last CBCT images for auto-contouring. Auto-plans were generated on last CBCT (PTV = bladder + 6 mm isotropic expansion except superiorly 3 mm, excluding bowel); 7-beam IMRT; 55 Gy in 20 fractions) via Eclipse Scripting API. Conventional plans were made on simulation CTs (PTV=bladder+1cm), transferred to last CBCT images based on the actual therapist registrations, and forward calculated dose. The quality of auto-contours was evaluated against manual contours. Plans on last CBCT images were evaluated with RTOG 0926 constraints as reference: PTV V100% > 95%, bladder (CTV) V100% > 100%, bowel D0.03cc<54–55Gy, rectum V55Gy < 50%. Summary statistics (mean ± SD) were computed for ART and conventional plans. Paired comparisons (ART vs conventional per patient) were tested with Wilcoxon signed-rank test.

Results:

The target bladder coverage was 100% ± 0% (ART) versus 93.5% ± 24.8% (conventional). Bowel max dose (D0.03cc) was 51.8 ± 3.9Gy (ART) versus 66.8 ± 6.0Gy (conventional) (p < 0.001). No ART cases had rectum dose above 55Gy, whereas it was 2.9%±6.0% by conventional treatment (p=0.028). PTV coverage was 95.7% ± 0.9% (ART) versus 92.6% ± 24.6% (conventional) (p = 0.013). ART achieved significantly better target coverage and lower bowel and rectum doses. The agreement (Dice similarity coefficient) between manual and auto contours were 93±2% for bladder and 86±4% for rectum and bowel combined. Auto-contour and auto-plan runtimes were approximately 0.5 and 1.5 minutes, respectively.

Conclusion:

Online bladder ART using DL-based auto-contouring and auto-planning on CBCT is feasible and yielded superior target coverage and reduced bowel and rectum dose compared with conventional planning in this cohort. These dosimetric results support further development of online ART for bladder cancer to meet RTOG 0926 objectives and potentially reduce adverse events.