Early Prediction Model of Radiation-Induced Xerostomia Based on Radiomics during Radiotherapy for Nasopharyngeal Carcinoma

Purpose/Objective(s)

Radiation-induced xerostomia remains one of common side effects for radiotherapy (RT) of nasopharyngeal carcinoma (NPC) and severely affects the quality of life. Early prediction of xerostomia is extremely important, but few studies mention it. We aim to build the early prediction model of xerostomia based on radiomics during RT by artificial intelligence for NPC patients.

Materials/Methods

Seventy-five pathologically confirmed squamous NPC patients were prospectively enrolled in our institution. All patients received a total radiation dose of 66 to 70Gy in 33 fractions at 5 fractions per week using Image-guided intensity modulated radiotherapy (IGRT). The patients received 3-5 cycles of cisplatin-based chemotherapy. MR images with T2-weighted (T2WI), contrast-enhanced T1-weighted (CE-T1WI) and diffusion-weighted imaging (DWI) (b = 0, 500, 800 s/mm²) were acquired on 3.0T MRI. RTOG criteria was used to evaluate xerostomia at 3 months after the completion of radiation. Delivered radiation dose of bilateral parotid and submandibular glands, MR images were collected at pre-RT, 5^th, 15^th fractions and immediately post-RT, respectively. Bilateral parotid and submandibular glands were delineated as the regions of interest (ROI) on CE-T1WI, T2WI, and DWI. We extracted a variety of features from all the images, and selected the optimal features combination by variance analysis. We built the prediction model by support vector machine (SVM). Receiver-operating characteristic (ROC) curve was used to evaluate predictive performance.

Results

The average age was 44.88 years (23-79 years). 24 were female and 51 were male. The mean ADC values of parotid and submandibular glands increased gradually during radiotherapy. Furthermore, the mean ADC values of salivary glands began to increase at 5^th fractions with minimal volume change. For xerostomia, 34 patients with G3-5 were categorized as toxicity cases, and 41 patients with G0-2 were classified as non-toxicity cases. Based on radiomics, clinical characteristics (sex, age, 8^th TNM staging) and radiation dose, we built 15 prediction models. Among them, the best prediction model consists of radiation dose, clinical parameters and ΔRF (the delta-radiomics features from pre-RT to 5^th fractions) which includes 6 ipsilateral parotid features, 8 contralateral parotid features, 7 ipsilateral submandibular features, and 7 contralateral submandibular features. The area under ROC curve is 0.7641, with a precision of 0.7444, a sensitivity of 0.7238, and a specificity of 0.8424.

Conclusion

Early prediction model including delta-radiomics, radiation dose, and clinical parameters exhibits superior sensitivity and specificity. This is an effective model for predicting xerostomia during the early stage of RT, and the clinicians could use it to adjust the treatment strategy and reduce the damage to salivary glands.