Interpretable Deep Learning Enables MRI-Based Virtual Biopsy for Molecular Stratification of Diffuse Gliomas

Background: The integration of molecular profiling in the classification of gliomas is essential to inform on patients’ prognosis and guide treatment. Radiomics and Deep Learning (DL) algorithms proved cost-effective non-invasive alternatives to predict genetic alterations. However, these methodologies have limitations due to tumor heterogeneity, variability in imaging acquisition protocols, the use of tumor segmentation masks and the lack of external validation cohorts. In this study, we developed a DL-based methodology to achieve a robust molecular stratification of adult gliomas based on whole-brain 3D diagnostic, structural MRI scans. Methods: Pre-operative T1-weighted before and after gadolinium administration, T2-weighted and T2-weighted fluid-attenuated inversion recovery sequences and molecular information (IDH and 1p19q mutation status) of 870 gliomas across 14 centers were collected. Two independent DL models were implemented, optimized, and subsequently integrated into a tiered framework for classifying tumors as IDH-WT, IDH-mutant 1p/19q intact and IDH-mutant / 1p/19q co-deleted. The final classifier was tested on 227 patients across 4 centers, which were not included in the training phase. Results: In the external test set, the tiered model achieved an AUC of 0.93 [IDH-WT], 0.91 [IDH-mutant 1p/19q intact] and 0.89 [IDH-mutant 1p/19q co-deleted], improving on previous approaches. Generated attention maps of enhancing tumor regions identified IDH status; maps of non-enhancing regions distinguished 1p/19q intact from co-deleted tumors. Conclusions: We demonstrated that the tired DL model developed in this study is effective to stratify gliomas according to their IDH and 1p/19q status, relying on standard 3D MRI scans from multicenter validation cohort. This strategy has the potential to be implemented in clinical practice to predict a broader range of mutations and epigenetic alterations and avoid an invasive procedure in patients with comorbidities or tumors in eloquent areas. The application of the tired DL model will be instrumental to implement clinical trials testing the efficacy and safety of neoadjuvant therapy in patients with glioma.