Deep learning-based spatial tumor phenotyping for radiogenomic prediction of the IDH-genotype in preoperative glioma

Motivation: The isocitrate dehydrogenase (IDH)-genotype is a central determinant for the diagnosis, therapy and management of patients with adult-type diffuse glioma. Goal(s): To investigate spatial tumor characteristics using an U-Net-based CNN and to evaluate its diagnostic potential for predictive IDH-genotyping. Approach: Participants with treatment-naïve glioma were investigated using an U-Net-based CNN based on a standard MRI protocol computing a compartmental tumor mass (CTM)-ratio. Results: A total of 207 patients (IDH-mutant, 72, IDH wild-type, 135) with treatment-naïve glioma were evaluated. IDH-mutated glioma presented greater CTM-ratio (P<.001) unlike IDH wild-type glioblastoma (P<.001) - which determined the IDH-genotype with excellent diagnostic performance (AUC=0.85, P<.001). Impact: We performed DL-based spatial mapping of treatment-naïve glioma and computed the CTM-ratio for the evaluation of spatial tumor characteristics, which enabled robust classification of the IDH-genotype - with implications for the clinical management of adult-type glioma.