Development of programs to predict the occurrence of mucositis from digital imaging and communications in medicine data by machine learning in head and neck volumetric modulated radiotherapy

Abstract

AbstractVolumetric modulated arc therapy (VMAT) with cisplatin for head and neck cancer is often accompanied by symptoms of pharyngeal and oral mucositis. However, no standard medical program exists for the prevention and treatment of mucositis, and the mechanisms of mucositis have not yet been fully proven. Therefore, adaptive radiotherapy (ART), which is a re‐planning process, is administered when severe mucositis develops during the treatment period. We extracted the treatment plans of patients who developed severe mucositis from DICOM data and used machine learning to determine its quantitative features. This study aimed to develop a machine learning program that can predict the development of mucositis requiring ART. This study included 61 patients who received concurrent chemotherapy and radiotherapy (RT). For each patient, the equivalent square field size of each segmental irradiation field used for VMAT, dose per segment (Gy), clinical target volume high, and mean dose of the oral cavity (Gy) were calculated. Furthermore, 671 five‐dimensional lists were generated from the acquired data. Support vector machine (SVM) and K‐nearest neighbor (KNN) were used for machine learning. For the accuracy score, the test size was varied from 10% to 90%, and the random number of data extracted in each test size was further varied from 1 to 100 to calculate a mean accuracy score. The mean accuracy scores of SVM and KNN were 0.981 ± 0.020 and 0.972 ± 0.033, respectively. The presence or absence of ART for mucositis was classified with high accuracy. The classification of the five‐dimensional list was implemented with high accuracy, and a program was constructed to predict the onset of mucositis requiring ART before treatment began. This study suggests that it may support preventive measures against mucositis and the completion of RT without having to re‐plan.