M2-YOLO: Multi-Task Instance Segmentation and Non-Invasive Molecular Profiling of Gliomas via Preoperative MRI

Gliomas are the most common malignant primary brain tumors. The 2021 WHO classification requires the integration of molecular biomarkers, including IDH mutation, MGMT promoter methylation, and 1p/19q codeletion, alongside histological criteria, yet their assessment still relies on invasive surgical biopsy. We present M2-YOLO (Multi-task Molecular YOLO), a unified deep learning framework that simultaneously performs glioma instance segmentation and non-invasive molecular biomarker prediction from preoperative multi-sequence MRI. The proposed approach was evaluated on the multi-site UTSW-Glioma dataset comprising 625 patients. M2-YOLO achieved Dice scores of 0.853 for edema and 0.697 for tumor core segmentation, together with AUC-ROC values of 0.849 for IDH mutation and 0.718 for 1p/19q codeletion prediction. MGMT prediction yielded an AUC of 0.520, consistent with the limited discriminative capacity of anatomical MRI for this biomarker. These results demonstrate the potential of a unified multi-task framework for combining tumor delineation and molecular characterization within a single non-invasive workflow, supporting future applications in precision neuro-oncology.