KEY POINTS
- This retrospective single-center study included 100 consecutive patients with non-metastatic nasopharyngeal carcinoma treated with definitive IMRT from 2022–2024. All tumors were left-dominant to standardize the geometric relationship between the target and ipsilateral parotid; 67% had stage III and 22% stage IV disease.
- Radiotherapy delivered 70 Gy to PGTVnx, 68 Gy to nodal gross disease, 66 Gy to PTV1, 60 Gy to PTV2 and 54 Gy to PTV3 in 32 fractions. Fourteen CT-derived shape features were extracted specifically from the spatial overlap between PTV2 and the left parotid, then combined with left-parotid V20, V30, V50 and mean dose.
- Xerostomia was assessed at a median 3.1 months after RT: 53% had grade 1, 42% grade 2 and 5% grade 3, making 47% of the cohort the grade ≥2 prediction class. A random forest classifier was evaluated through 100 stratified 80:20 train-test splits with feature selection performed within each training set.
- The combined radiomic-dosimetric model achieved mean test AUC 0.78 ± 0.09 and accuracy 0.71 ± 0.09, with sensitivity 0.64 and specificity 0.77. Its Brier score was 0.195, with mean calibration slope 0.938 and intercept 0.023.
- Adding shape information produced a clear improvement over dosimetry alone: mean AUC increased from 0.66 to 0.78, accuracy from 0.61 to 0.71, specificity from 0.64 to 0.77, and Brier score improved from 0.253 to 0.195.
- V30 was the most influential individual feature, followed by overlap-region MeshVolume and Sphericity. Shape features occupied 8 of the 10 highest importance positions, while Elongation, Maximum2DDiameterSlice, Flatness, LeastAxisLength and MajorAxisLength were selected in 100% of the 100 iterations alongside V20.
- Generalizability is the central limitation. Mean training AUC was 0.984 versus 0.78 on held-out testing, indicating appreciable overfitting; there was no independent external cohort, only left-dominant tumors were included, clinical factors were omitted, and the study did not compare the overlap-region approach directly with whole-parotid radiomics.
CLINICAL TAKEAWAY
The geometry of how much and in what shape the elective target overlaps the parotid may contain toxicity information beyond conventional dose-volume metrics. The improvement over a dosimetry-only model is interesting, but the substantial training-test gap and lack of external validation make this a proof-of-concept rather than a clinically usable xerostomia predictor.