A comprehensive machine-learning model applied to MRI to classify germinomas of the pineal region

Background: Pineal region tumors (PRTs) are highly histologically heterogeneous. Germinoma is the most common PRT and is treatable with radiotherapy and chemotherapy. A non-invasive system that helps identify germinoma in the pineal region could reduce lab exams and traumatic therapies.Methods: In this retrospective study, 122 patients with histologically confirmed PRTs and pre-operative multi -modal MR images were included. Radiomics features were extracted from different ROIs and image sequences separately. A computational framework that combines a few classification and feature selection algorithms were used to predict histology with radiomics features and demographics. We systemically benchmarked performance of models with feature matrices from all possible combinations of ROIs and image sequences. The Area under the ROC Curve (AUC) was then used to evaluate model performance.Results: Models with demographics and radiomics features outperform radiomics-only or demographics-only models. The best demographical-radiomics model reached the highest AUC of 0.88 (CI95%: 0.81-0.96). Through the comprehensive evaluation of possible sequence combinations in the differential diagnosis of pineal tumor, T1 and T2 emerged as the most informative sequences for the task. There is imbalanced usage of feature classes as we analyze their proportion in all models.Conclusions: The demographical-radiomics model can accurately and efficiently identify germinomas in the pi-neal region. The preference for MRI sequences, radiomics feature classes, features selection and classification algorithms provide a valuable reference for future attempts at developing classifiers on medical images.