Delta-radiomics features during radiotherapy improve the prediction of late xerostomia.

The response of the major salivary glands, the parotid glands, to radiation dose is patient-specific. This study was designed to investigate whether parotid gland changes seen in weekly CT during treatment, quantified by delta-radiomics features (Δfeatures), could improve the prediction of moderate-to-severe xerostomia at 12 months after radiotherapy (Xer12m). Parotid gland Δfeatures were extracted from in total 68 planning and 340 weekly CTs, representing geometric, intensity and texture characteristics. Bootstrapped forward variable selection was performed to identify the best predictors of Xer12m. The predictive contribution of the resulting Δfeatures to a pre-treatment reference model, based on contralateral parotid gland mean dose and baseline xerostomia scores (Xerbaseline) only, was evaluated. Xer12m was reported by 26 (38%) of the 68 patients included. The most predictive Δfeature was the contralateral parotid gland surface change, which was significantly associated with Xer12m for all weeks (p < 0.04), but performed best for week 3 (ΔPG-surfacew3; p < 0.001). Moreover, ∆PG-surfacew3 showed a significant predictive contribution in addition to the pre-treatment reference model (likelihood-ratio test; p = 0.003), resulting in a significantly better model performance (AUCtrain = 0.92; AUCtest = 0.93) compared to that of the pre-treatment model (AUCtrain = 0.82; AUCtest = 0.82). These results suggest that mid-treatment parotid gland changes substantially improve the prediction of late radiation-induced xerostomia.