A Predictive model of radiation-related fibrosis based on radiomic features of Magnetic Resonance Imaging

Abstract

Radiation-related fibrosis of normal tissues remains unpredictable. To predict the fibrotic level of neck muscles after radiotherapy, we established a predictive model by using the magnetic resonance imaging (MRI) before and after radiotherapy based on Radionics in nasopharyngeal carcinoma patients. A total of three hundred and seven patients who received radical radiotherapy at a single institution were enrolled in this study. Patient's clinical were recorded systematically. The median follow-up time is 18 months (range, 11-22 months). According to the symptoms and physical signs, all patients were divided into two groups, the mild fibrosis group, and the severe fibrosis group. All MRI images included three sequences: T1 weighted scans, T2 weighted scans, and contrast-enhanced T1 weighted scans. Regions of interests (ROI), were delineated manually on an open infrastructure software platform, including sternocleidomastoids (SCMs), trapezius (T), levator scapulae (LS), and scalenus (S) muscles. The ROIs were divided into two parts by the level of the cricoid cartilage. The dose distribution was matched to each ROI individually via a software program, Min, Max, and Mean dose parameters were collected afterwards. One hundred ninety features were extracted via the software platform from pre- and post-radiotherapy images. The magnitude of change of each feature was computed as ‘delta change= [post-pre]/pre and then correlated to the grade of neck fibrosis using a machine learning algorithm, XGBoost. First, the image features delta changes with label information were both divided into training and testing sets. Then, based on the training dataset, a well-trained prediction model was provided by XGBoost. Afterward, the predictive model performance was assessed on the ‘testing’ set and reported in terms of area under the receiver operating characteristic curve (AUC). Most of the patients enrolled are male (73.6%), mean age was 47 years, locally advanced stage (89.3%), receiving concurrent chemo-radiotherapy as the primary treatment (90.6%). By the end of the final follow-up, most of the patients were rated as mild fibrosis (85.6%). We found that the image features from T1+C modality outperformed all the other feature combinations in terms of the prediction accuracy, with an AUC of 0.61. The prediction model based on radiomics features from T1+C modality have the best performance in prediction of the grade of severity of post-radiotherapy neck fibrosis. This might help to guide radiotherapy treatment planning and organs at risk sparing for the better quality of life outcomes.Abstract 3389; Table 1The categorization of muscle fibrosis on the neck after treatment among 307 patientsLevelConditionsMild1. Self-Rating: 0-4 points2. No facial edema3. No upper limb pains4. Neck activity is unrestrictedSevere1. Self-Rating: 5-10 points2. Facial edema3. Upper limb pains4. Neck activity is limited Open table in a new tab