2615 - Promise and Pitfalls of AI-Generated Contours of the LAD Coronary Artery in Thoracic Radiotherapy

Purpose/Objective(s): Radiation dose to the left anterior descending (LAD) coronary artery is associated with major adverse cardiac events and all-cause mortality in locally advanced non-small cell lung cancer (NSCLC), emphasizing the importance of contouring cardiac substructures. Our group performed a dosimetric comparison of LAD contours between a validated deep-learning-based automated segmentation tool and a physician to probe the limits of our artificial intelligence (AI) model’s performance on 4D-CTs.

Materials/Methods: An IRB-approved retrospective review of twenty patients who received notable V15 Gy LAD dose during 5-fraction stereotactic body radiotherapy (SBRT) or palliative RT at our institution was conducted. A physician manually contoured the LAD_PRV. Then, a proprietary 3D U-Net deep learning model created LAD_REGION contours, running along the interventricular groove occupied by the LAD. Inclusion ratios (IR) reported the percentage of the manual contour within the AI-generated contour to assess geometric accuracy. AI failure criteria included discontinuous contours causing an IR below 0.4 or over-contoured regions with a volume over 0.5 cm³. After removing AI failures, linear regression of dose-volume histogram (DVH) metrics (D1%, mean dose, V10Gy, and V15Gy) was performed and coefficient of determination (R²) and mean absolute error (MAE) were calculated to understand dosimetric accuracy.

Results: Of the twenty patients, six received palliative RT and 14 received SBRT. Four AI failures were identified. Failures were outlying contours where abnormalities (pleural effusion, tumor, cardiac lead) adjacent to the heart caused AI to misidentify the LAD. Table 1 compares DVH metrics by R² and MAE for the other 16 patients. Mean and median IRs were 0.72 and 0.75, respectively, with a standard deviation of 0.14.

Conclusion: An AI-segmented LAD_REGION can strongly correlate with a manual LAD_PRV, but interpreting abnormalities remains a challenge for the AI model. Contouring errors can be corrected with clinical judgment. Refining this AI model to delineate the LAD will enable large-scale analyses correlating dose metrics with oncologic outcomes and toxicity.