3752 - Evaluation of AI-Generated Contours for MRI-Based Gynecological HDR Brachytherapy

Purpose/Objective(s):

Magnetic resonance image guided adaptive HDR brachytherapy (MR-IGABT) is the gold standard for locally advanced cervical cancer. EMBRACE/GEC-ESTRO guidance requires repeated MR imaging with the applicator in situ for brachytherapy (BT) planning. Although AI-based auto-contouring is widely used in external beam radiation therapy (EBRT) to streamline planning workflow, improve efficiency and reduce inter-observer variability, its use in BT remains limited due to training data challenges (applicators, artifacts, and organ deformation). Given the resource-intensive and time-consuming nature of MR-IGABT, we evaluated a commercial deep-learning auto-contouring solution specifically trained for MRI-based gynecologic (GYN) HDR BT and assessed its suitability for routine integration into our MR-IGABT workflow.

Materials/Methods:

Ten consecutive patients treated for cervical cancer per EMBRACE/GEC-ESTRO concordant MR-IGABT using intracavitary applicators with or without interstitial needles were retrospectively analyzed. All patients received 45 Gy EBRT followed by four MR-IGABT fractions (total 28 Gy; 4 insertions each patient), yielding 40 fractions. The commercial auto-contouring software trained for MR-IGABT was applied to contour organs at risk (OARs: bladder, rectum, sigmoid). Using built-in timestamps, OAR contouring time per fraction was recorded for auto-contouring versus manual contouring. Quantitative agreement with clinically approved contours (manual contouring) was evaluated using the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and 95th-percentile HD (95% HD).

Results:

Mean (SD) DSCs were 0.80 (0.10) for bladder, 0.61 (0.14) for rectum, and 0.58 (0.23) for sigmoid. Mean (SD) HDs were 2.1 mm (1.1), 5.4 mm (2.7), and 6.7 mm (8.2), respectively; corresponding 95% HDs were 5.2 mm, 22 mm, and 21 mm. The bladder model performed best. One outlier patient had a sigmoid DSC of 0 (no overlap) and HD of 30 mm (the largest deviation). However, the sigmoid laid outside the treated region and did not impact dosimetry (D2cc absolute difference < 0.1 Gy). Excluding this outlier improved sigmoid performance to DSC 0.65 (vs. 0.58 overall) and HD 4.0 mm (vs. 6.7 mm overall). Auto-contouring significantly reduced OAR contouring time: mean 5.5 min (range 4.0–7.0) versus 22 min (range 18–25) for manual contouring (p < 0.001).

Conclusion:

A commercially available MRI GYN BT trained auto-contouring solution demonstrated good agreement with clinically approved OAR contours and yielded statistically significant reductions in contouring time for MR-IGABT. These results support its feasibility for routine adoption to enhance MR-IGABT workflow efficiency and consistency.