Performances of AI-based automatic delineation of CTV and Organs-at-Risk for breast radiotherapy planning: including inter-observer variability in the evaluation process.

PURPOSE: This study evaluated the performance of three AI-based segmentation tools for breast clinical target volumes (CTV) and organs at risk (OARs): a commercial solution, an open-source tool, and an in-house UNet model. Inter-observer variability (IOV) was computed and used as a benchmark for the subsequent considerations on the automatic contours. METHODS: Forthy breast cancer patients were analyzed. Four physicians, including a resident, manually contoured the breast CTV and OARs (heart and contralateral breast). Physicians reviewed and edited the contours generated by the commercial AI-based tool and assigned quality Similarity. Geometric accuracy was assessed using Dice Similarity Coefficient, Hausdorff Distance (HD), 95th percentile HD, and Average Surface Distance. Treatment plans dosimetric differences were also analyzed. RESULTS: Results showed that automatic segmentation of the heart and contralateral breast achieved good agreement with manual contours and fell within the range of IOV for all three tools. However, regarding the commercial tool, for the breast CTV, discrepancies between manual and automatic contours were significantly greater than IOV (p < 0.05). The mean difference between automatic and edited contours was minimal for the contralateral breast (-0.02 Gy; SD: ±0.06 Gy right, ±0.12 Gy left) and for the heart in left-sided cases (-0.08 Gy; ±0.14 Gy). Editing the breast CTV required approximately 7 min, longer than manual contouring, suggesting no time-saving benefit. CONCLUSIONS: Commercial AI-based segmentation tools may reliably replace manual contouring just for select OARs, while the trained in-house UNet model reaches the IOV derived by this paper for CTV breast segmentation.