Semi-supervised learning for non-invasive radiopathomic mapping of treatment naïve glioma with multi-parametric MRI

Motivation: Radiopathomic mapping of glioma could improve standard of care by helping guide surgical resection and subsequent treatment. Most current methods for predicting tumor pathology using MRI neglect intra-tumoral heterogeneity. Goal(s): We aim to use multi-parametric MRI and deep learning to spatially map pathology for treatment naïve glioma. Approach: We utilized histopathologically analyzed tissue samples taken during surgical resection with known coordinates on pre-surgical multi-parametric MRI and semi-supervised ensemble networks. Results: Our model classifies Ki-67 with an AUROC of 0.84 and 0.73 for combined Ki-67 and percent cancerous cells. Including physiologic MRI and pretraining on patches of unknown pathology improved performance. Impact: We performed radiopathomic mapping in patients with newly-diagnosed glioma using presurgical physiological + anatomical MRI and semi-supervised ensemble networks and achieved AUROCs of 0.84 and 0.73 for Ki-67 and combined Ki-67 and % cancerous cells, respectively.