Non-invasive identification of clinically actionable gene fusions in gliomas using deep learning-enabled analysis of magnetic resonance imaging scans

Abstract

Gliomas are the most common primary malignant brain tumors in adults, with prognosis varying substantially across molecular subtypes. Oncogenic gene fusions have recently emerged as a promising class of actionable therapeutic targets, yet their detection relies on invasive tissue sampling via biopsy or surgical resection. Here we present a deep learning framework that leverages magnetic resonance imaging (MRI) to non-invasively predict gene fusion status in gliomas. Using contrastive self-supervised learning, we pretrained a foundation model on a diverse cohort of 30,270 brain MRI images from 10,336 patients across multiple institutions and tumor types. Validation on brain tumor type classification showed that the model learned meaningful representations (AUC: 0.84), outperforming existing foundation models. Application to gene fusion status classification, however, yielded more limited discrimination (AUC: 0.54), suggesting that gene fusions may manifest as comparatively subtle or heterogeneous MRI patterns. Our findings demonstrate that foundation models can learn transferable representations across diverse neuroimaging tasks, although non-invasive identification of gene fusions in gliomas remains an open challenge.