Evaluation of Deep Learning-Based Auto-Segmentation of Organs-at-Risk for Breast Cancer Radiation Therapy

Purpose/Objective(s) A large inter-physician variation can be seen when delineating organ-at-risks (OARs) for breast radiotherapy. This study aimed to externally validate the performance of deep learning-based auto-contouring system (ACS) for breast radiotherapy. Materials/Methods Eleven experts from two institutions were asked to delineate 9 OARs of 10 cases of breast radiotherapy. Then, auto-contours were provided to the experts for correction. To compare the performance of auto-, corrected-auto-, and experts’ manual contours, Dice similarity coefficient (DSC) and Hausdorff distance (HD) between each contour and the best manual contour were used, where higher DSC and lower HD means better geometric overlap. Results Total mean time for 9 OARs was 37 ± 20 min for manual and 6 ± 5 min for corrected-auto-contours. Among the DSC of experts’ manual contours and an auto-contour, DSC of an auto-contour ranked the second place and HD ranked the first place. Better performance was shown in corrected-auto-contours than in manual contours (median DSC: 0.90 vs. 0.88; median HD: 4.5 vs. 6.5 mm). The inter-physician variations among experts were reduced in corrected-auto-contours, compared with manual contours (DSC range: 0.86–0.90 vs. 0.89–0.90; HD range: 5.1–9.1 mm vs. 4.3–5.7 mm). Among manual OARs, breast contours had the largest variations, which were most significantly improved with an aid of ACS. Conclusion ACS showed at least similar performance in OARs compared with experts’ manual contouring, which anticipates further applications of ACS to target volumes. ACS can be a valuable tool for improving the quality of breast radiotherapy and reducing inter-physician variability in clinical practice.