The deep topology of glioma

Motivation: Heterogeneity in neuro-oncology is a major barrier to clinical and academic innovation, resolvable only with expressive computational modelling, large-scale data, and high-performance computing. Goal(s): To develop a comprehensive mathematical modelling framework illuminative to glioma's multi-modal heterogeneity. Approach: In the largest and most detailed international investigation of glioma patients (n=4908 across six different cohort sites), we undertake cartography to characterise the interplay between lesion development and the space they occupy. Results: We reveal the spatial properties of glioma, a deep auto-encoder representation, and from which we develop machine models predicting patient-personalised diagnosis and prognosis, disclosing the joint relationship between imaging appearances and underlying neuropathology. Impact: These works illustrate the benefit of computational modelling across clinical neuro-oncological imaging data in patient-personalised care, including diagnostic and outcome prediction, paving the way for future research and clinical translation.