Integration of site-specific deep learning models for automated OAR segmentation in craniospinal proton therapy: a geometric and dosimetric analysis.

PURPOSE: This study evaluated the feasibility of integrating RayStation deep learning auto-segmentation (DLS) models-originally trained for adult head and neck (HN), thorax-abdomen (TA), and male pelvis (MP) regions-for automated organs-at-risk (OARs) delineation in craniospinal irradiation (CSI) with intensity-modulated-proton-therapy (IMPT), focusing on geometric accuracy, dosimetric impact, and clinical efficiency.

METHODS

Forty patients (aged 2-28 years) with CNS embryonal tumors were retrospectively analyzed. Their planning CT datasets were sequentially processed through the HN, TA, and MP DLS models to auto-segment 60 OARs. Thirty-one OARs were compared with expert contours using Dice Similarity Coefficient (DSC) and Hausdorff Distance (HD). Spearman's rank correlation was used to examine associations between geometric metrics and patient age, BMI, and craniospinal length. Dosimetric evaluation was performed by recalculating IMPT plans on DLS-derived OARs.

RESULTS

The mean auto-segmentation time was 4.9 ± 1.0 min per patient. Across 31 OARs, mean DSC and HD were 0.73 and 2.45 mm, with 41.9 % and 93.5 % meeting good geometric criteria (DSC > 0.8, HD < 4 mm). Geometric accuracy showed significant correlations (p < 0.05) with age, BMI, and craniospinal length for up to 11 OARs. Accuracy decreased in children (<12 years), especially for the thyroid, bowel, and pelvic structures. Mean and maximum dose deviations were within 105 and 190 cGy(RBE), except for the mandible, esophagus, and anorectum.

CONCLUSION

The integrated DLS framework achieved reliable geometric and dosimetric performance across most OARs with substantial efficiency gains, offering a practical solution for rapid, standardization of CSI planning workflows under expert review.