2979 - Automated Delineation of the Parotid Stem Cell-Rich Region for Xerostomia-Sparing Radiotherapy

Purpose/Objective(s): Dose to the parotid stem cell–rich region (SCR), a geometrically defined parotid subregion, has been associated with xerostomia and evaluated in SCR-sparing radiotherapy studies. We trained an automated SCR segmentation model using the published geometric definition and evaluated its concordance with physician contours.

Materials/Methods: Three radiation oncologists and one medical physicist contoured the SCR on 100 head and neck planning CT datasets according to the published geometric description to train a deep-learning auto-segmentation model. For evaluation, five additional CT cases were contoured by five physicians using identical instructions. For each gland, physician interobserver Dice similarity coefficient (DSC) was calculated pairwise, and automated contours were compared with each physician contour. For each case and side, median physician–physician and physician–AI DSC were computed. Centroid displacement was defined as the Euclidean distance between the AI centroid and the physician consensus centroid (coordinate-wise median). Volume ratios (AI/physician median volume) assessed systematic bias. Physician–AI and physician–physician DSC were compared using Wilcoxon signed-rank testing across 10 SCRs.

Results: Physician interobserver agreement was high, with median physician–physician DSC 0.820–0.918 (left) and 0.813–0.926 (right). Median physician–AI DSC ranged 0.404–0.896 (left) and 0.352–0.872 (right). Median AI centroid displacement was 0.423 cm (left) and 0.388 cm (right), with most glands =0.5 cm. One case demonstrated substantially greater displacement with reduced overlap and marked under-segmentation. Median AI-to-physician volume ratio was 1.024 (left) and 0.976 (right), without consistent systematic bias. Physician–AI DSC was significantly lower than physician interobserver DSC (Wilcoxon signed-rank p=0.004).

Conclusion: The geometric SCR construct demonstrated strong physician reproducibility. An automated model trained on 100 datasets achieved generally favorable agreement, with one outlier case demonstrating substantial deviation. Automated SCR delineation appears feasible with physician oversight and enables scalable dosimetric evaluation of SCR-sparing strategies. Continued validation is warranted to support broader clinical implementation.