Room: Exhibit Hall | Forum 4
Purpose: We evaluated the usefulness of 3D Convolutional Neural Network (3D-CNN) for acute xerostomia prediction during the radiotherapy (RT) for Nasopharyngeal Cancer (NPC).
Methods: 3D-CNN is an end-to-end framework, reducing the difficulty of multi-stage processing of the prevalent radiomics methods. We proposal a 3D-CNN adapted from ResNet which can automatically extract efficient deep-level non-linear features. 3D-ResNet reduces reliance on the manual operation that requires professional domain knowledge.89 NPC patients were selected from a clinical trial which was registered on the clinicaltrials.gov (ID: NCT01762514). All patients received radical treatment based on IMRT with a prescription dose of 68.1 Gy in 30 fractions. Acute xerostomia was evaluated based on RTOG acute toxicity scoring (TS). The patients ages ranged 24â€“72y and each patient had five CT sets acquired in treatment position at 0áµ—Ê°, 10áµ—Ê°, 20áµ—Ê°, 30áµ—Ê° fractions during RT, and at 3-month after RT. Each CT volume contained 90~130 slices with 512x512 pixels. Pre-processing includes redundant slices reduction, position calibration, pixels HU value clip according soft-tissue threshold and non-relevant area filter. The 3D-ResNet model is implemented via Pytorch and all experiments are conducted on a server with 4 GeForce_GTX_Titan_Xp_GPU. Three experiments sets are conducted: TS0F to predict TS10F (TS0F->10F), TS0F to predict TS30F (TS0F->30F) and TS10F to predict TS30F (TS10F->30F). The fine-grained classification accuracy (FA) of four xerostomia levels were used to evaluate and predict the acute xerostomia level at 10F and 30F.
Results: Similar accuracy in three sets of experiments is observed. Each set of the experiment are conducted under a 5 fold cross-validation setting and the average accuracy are taken as the final judging metric. The model predicts the level of acute xerostomia with a FA of 98.82%, 97.50%, 98.75% on TS0F->10F,TS0F->30F and TS10F->30F, respectively.
Conclusion: This study demonstrates that radiation-induced acute xerostomia level can be predicted using 3D-CNN.
Funding Support, Disclosures, and Conflict of Interest: Student's Platform for Innovation and Entrepreneurship Training Program - 201813902075; 201813902071; 201713902050; Hubei Key Laboratory of Medical Information Analysis and Tumor Diagnosis and Treatment - PJS140011504; Natural Science Foundation of Guangdong Province - 2017A030310217; Pearl River S&T Nova Program of Guangzhou - 201710010162