2530 - Evaluation of an AI-Auto-Contouring Device with Clinical Contour Integration Capabilities, for Rectal Cancer Treatment Planning

Purpose/Objective(s): Accurate and consistent target delineation is essential for high-quality radiotherapy planning, particularly in rectal cancer where variability in contouring can significantly affect treatment outcomes. Artificial intelligence (AI)-based auto-contouring tools have shown promise in improving efficiency and reproducibility. This study evaluates the performance and usability of a commercial AI-assisted auto-contouring device, featuring customizable post-processing tools and automated integration of manually defined gross tumour volumes (GTVs) into AI-generated predictions.

Materials/Methods: 5/ planned 10 anonymized patients with rectal cancer CT simulation scans with expert-defined manual GTV and clinical target volume (CTV) contours were used to test the device. The AI model generates contours according to the Valentini guidelines for rectal cancer delineation. The reference manual CTV contours were created following the Royal College of Radiologists (RCR) consensus recommendations, as used in daily practice. The device automatically incorporated the provided manual GTV structures into the to-be generated AI-segmentations, and adjusted the AI generated LN contours into RCR compliant elective CTV ROIs (combining individual LN levels and addition of GTV+margin). Additionally, the bowel bag ROI was post-processed and compared with the manual contours, as the original AI structure differed significantly from the clinical contours. Quantitative performance was evaluated using standard geometric metrics (Dice similarity coefficient, Hausdorff distance).

Results: The AI-generated contours demonstrated good agreement with reference manual CTVs with an average Dice score of 0.87 and HD95 of 8.3mm. The GTV+margin scored an average Dice of 0.95 and HD95 of 7.5mm. The adapted bowel bag reached an average Dice score of 0.88 and HD95 of 7.9 mm. According to the statistical Sign test, the median Dice score was significantly higher than 0.85 (p < 0.05). The automated GTV integration was successful in all cases, with no need for additional manual steps for this purpose. Customizable post-processing features allowed for automatic contour refinement.

Conclusion: Our evaluation indicates that the tested AI-based auto-contouring device can accurately and efficiently generate rectal cancer target structures aligned with established guidelines. The combination of automated GTV incorporation and user-adjustable post-processing enhances both the adaptability and clinical utility of the system. Our results will include all 10 cases at the time of the meeting. With further validation in larger datasets, this technology has the potential to streamline radiotherapy planning and reduce inter-observer variability in rectal cancer contouring.