Main Session
Sep
27
PQA 01 - Gastrointestinal Cancer and Central Nervous System
2009 - Clinical Acceptability of AI-Generated Target Contours for Anal and Rectal Cancer Radiotherapy
Presenter(s)
Frank Arturi, MEng - Memorial Sloan Kettering Cancer Center, New York, NY
F. J. Arturi1, S. Blum2, P. P. McCann2, E. K. Liu1, R. Ravella1, H. Veeraraghavan2, J. Jiang2, J. J. Cuaron1, C. H. Crane1, A. J. Wu1, M. T. McMillan1, R. A. Weber1, M. Zinovoy1, K. G. Hockemeyer1, A. L. Damato2, P. B. Romesser1, D. A. Roth O’Brien1, and S. Elguindi2; 1Department of Radiation Oncology, Memorial Sloan Kettering Cancer Center, New York, NY, 2Department of Medical Physics, Memorial Sloan Kettering Cancer Center, New York, NY
Purpose/Objective(s):
Accurate clinical target volume (CTV) delineation for anal and rectal cancer radiotherapy is time-intensive and subject to inter-observer variability. While AI auto-segmentation has shown great utility for organs-at-risk, validation of real-world clinical use of AI-generated target volumes remains limited. We hypothesized that CTV contours developed through an in-house AI platform would generate clinically acceptable contours. Dosimetric equivalence was assessed in an exploratory subset.Materials/Methods:
Using the Mentored AI Transformer Environment (MATE), an in-house software platform for AI segmentation model creation and deployment, radiation oncologists curated training contours for CTV-A (primary tumor, mesorectum, obturator, internal iliac and pre-sacral nodal regions; n=44) and CTV-ABC (CTV-A plus external iliac and inguinal nodal regions; n=31). For validation, AI contours from 30 cases (15 CTV-A, 15 CTV-ABC) were independently rated by 9 radiation oncologists using a 4-point clinical acceptability score (1=no edits, 4=clinically unacceptable; score =2 = clinically acceptable). Following validation, AI contouring was integrated into routine workflow for 29 patients (19 CTV-A, 10 CTV-ABC). Geometric agreement between AI-generated and physician-edited contours was quantified using Dice similarity coefficient (DSC) and 95th percentile Hausdorff distance (HD95). A subset (n=9) of CTV-A patients compared independently optimized plans using AI CTVs alone versus physician-edited contours across dosimetric metrics (D95%, V100%, V105%, Dmax, Dmin) using Wilcoxon signed-rank tests.Results:
Across 202 physician ratings, AI contours were clinically acceptable in 72% of CTV-A (mean score 2.38 ± 0.70) and 60% of CTV-ABC assessments (mean score 2.46 ± 0.60), with high inter-rater reliability (ICC = 0.81). In clinical deployment, geometric agreement between AI and final clinical contours was high: CTV-A DSC 0.89 ± 0.13; CTV-A HD95 2.2 ± 2.7 mm; CTV-ABC DSC 0.84 ± 0.05; CTV-ABC HD95 2.9 ± 2.7 mm. In an exploratory dosimetric subset (n=9), plans generated from AI contours showed no significant differences across all evaluated metrics (D95%: 100.4% vs. 100.4%, p=0.31, V100%: 97.4% vs. 99.1%, p=0.36, V105%: 0.6% vs. 0.2%, p>0.99, Dmax: 106.2% vs. 105.8%, p=0.48, Dmin: 97.1% vs. 97.8%, p=0.07).Conclusion:
Using MATE, AI auto-segmentation models for anal and rectal cancer CTVs were developed, validated, and clinically deployed. AI contours were clinically acceptable for most cases with strong geometric agreement and dosimetric equivalence to final clinical contours in a subset of CTV-A contours. Lower acceptability for CTV-ABC highlights opportunities for continued refinement through the same platform-based workflow.