Radiomics and dosiomics in radionecrosis prediction in brain metastasis treated with Stereotactic Radiation Therapy: a machine learning approach.

INTRODUCTION: Stereotactic Radiotherapy plays a main role in Brain Metastases treatment. Radiomics and Dosiomics, coupled with Machine Learning approaches are emerging in radiation oncology as support in clinical decision-making workflow. In this study a machine learning approach was used to predict the late toxicities induced by Stereotactic Radiotherapy including clinical, radiomics and dosiomics features extracted from patients. MATERIALS AND METHODS: Lesions contribution, concomitant and/or sequential treatments were investigated by the models. Data were then split in training and test datasets 70:30. Features selection process, based on selectFromModel and selectKBest (python library kit) and the balancing SMOTE algorithm were applied on training cohort to optimize eleven machine learning models. 37 patients were considered exhibiting 113 brain metastases. Radionecrosis occurred in 21 (18.6 %) brain metastases. Four datasets, 1) clinical patients' data, 2) clinical & radiomics data, 3) clinical & dosiomics data, 4) the combination of 2) and 3), were investigated by means of the machine learning models. RESULTS: The best valuable Receiver Operating Characteristics-Area Under the Curve (ROC-AUC) values were reported for all datasets: in 1) K-Nearest Neighbors yields 80 % (C.I 55-97 %), in 2) Extra Trees gives 80 % (C.I 61-94 %), in 3) Logistic Regression reaches 75 % (C.I 50-94 %) and in 4) Random Forest gives 76 % (C.I 58-91 %). All models presented patients' age at RT and primary cancer diagnosis as clinical features. CONCLUSION: This approach may positively impact on patients' quality of life, helping radiation oncologists to improve patient-specific trial, and to reduce severity of the radiotherapy-induced toxicities.