3004 - A Multi-Dimensional AI-Based Quality Control System for Simulation CT and Target Delineation In Cervical Cancer Radiotherapy

Purpose/Objective(s): This study presents the development and validation of a multidimensional, automated quality control system for cervical cancer radiotherapy, integrating AI-based assessment of simulation CT image and target delineation. The system enables objective, end-to-end evaluation to enhance treatment precision, consistency, and safety.

Materials/Methods:

We retrospectively analyzed 121 cervical cancer patients using an AI-driven automated QC pipeline comprising simulation CT image and target delineation modules. CT image QC included: scan range verification via bone segmentation; setup accuracy assessment using radiopaque markers and spinal canal fitting; immobilization evaluation via air-tissue ratio; bladder volume and rectal air measurement; contrast phase determination from iliac vein attenuation; and metal artifact detection using Vision Transformers. Delineation QC employed a four-dimensional outlier detection strategy: integrity check for spurious contours/holes; AI consistency comparing physician contours with nnU-Net auto-segmentations (DSC, HD95); geometric shape analysis (volume, sphericity, eccentricity); and spatial overlap assessment (CTV-OAR distances). A hybrid ensemble (Isolation Forest, SVM, LOF) identified anomalies using multidimensional features.

Results: The automated QC system successfully processed all 121 patients. Scan range compliance was 100%. Setup accuracy pass rates were high for marker alignment (96.5%), couch tilt (96.7%), and patient tilt (93.4%). Vacuum cushion immobilization pass rate was 84.3%. Bladder volume compliance (150–300 mL) was only 33.9%, and 83.5% of patients had rectal air >20%, indicating suboptimal organ preparation. Contrast enhancement was appropriate in 49.6% of scans; metal artifact detection achieved 100% accuracy (4/4 cases).For delineation QC, integrity check identified spurious points in 10.7% (13/121) and internal holes in 6.6% (8/121). Compared with AI auto-segmentations, mean Dice was 0.842±0.038 (HD95: 11.1±6.7 mm), with 10 consensus outliers mainly due to cystic structure inclusion. Geometry outliers (11 cases) showed cranial truncation. Spatial overlap analysis identified 11 cases with abnormally small CTV-OAR distances (bladder: 3.3–6.0 mm; rectum: 1.7–7.5 mm), raising target coverage concerns.

Conclusion: This study presents an AI-driven, multidimensional QC system for cervical cancer radiotherapy, enabling objective, end-to-end assessment from CT simulation to contouring. It transforms QA into a standardized, quantitative process and detects critical deviations such as poor preparation and under-coverage. The system offers a practical tool to improve precision and safety.