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

• DOI: 10.1016/j.compbiomed.2022.106366
• Publication Date: 2022-11-26T18:32:53Z
• Authors: Ningrong Ye; Qi Yang; Peikun Liu; Ziyan Chen; Xuejun Li
• Keywords: Machine Learning; 03 medical and health sciences; 0302 clinical medicine; ROC Curve; Humans; Germinoma; Magnetic Resonance Imaging; Retrospective Studies; 3. Good health

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