Objective: Alongside evaluating the geometric and dosimetric performance of an AI auto-segmentation algorithm on conventional (primary) organs at risk (OARs), the aim of this work was to highlight its added advantages in terms of the dosimetric evaluation of secondary OARs, which are commonly under-reported despite their influence on overall toxicity. Methods: The study included 50 head and neck cancer patients, with volumetric modulated arc radiotherapy (VMAT) plans created based on both manual contouring and auto-contouring using a commercially available auto-segmentation tool. Quantitative assessment of auto-contouring was undertaken on 10 primary OARs using geometric performance metrics (the Dice Similarity Coefficient (DSC), Hausdorff distance (HD), sensitivity, and precision) as well as dosimetric differences between manual vs. auto-segmentation. Once the algorithm was validated on primary OARs, a dosimetric assessment of 29 secondary OARs delineated by the tool was conducted. Results: Manual contouring required 21.20 ± 2.25 min, while 9.25 ± 1.42 min were needed for auto-segmentation with adjustment. Dosimetric differences were found for the brainstem and the right submandibular gland. The mandible presented the maximum HD value of 24.09 ± 18.83 mm. Sensitivity, precision and DSC ranged from 0.77 to 0.93. For the 29 secondary OARs, 1500 dosimetric values were collected. Of these, 8.5% exceeded the dose constraints. The most impacted OAR was the constrictor muscle, exceeding the constraint in 56% of cases, while the glottis came in second. The dosimetric results also raised concerns regarding organs without defined dose constraints. Conclusions: When validated for primary OARs, auto-segmentation offers a more comprehensive dosimetric evaluation of secondary OARs to further reduce the dose to sensitive structures.
Costin et al. (2026) studied this question.