Deep learning classifier for histologic and molecular subtypes of gliomas and glioneuronal/neuronal tumors on multiparametric MRI

A deep learning MRI system distinguished six glioma/glioneuronal tumor subtypes with external test AUCs 0.69–0.92 and accuracy 0.81–0.95 in 1,844 patients, best for pilocytic astrocytoma versus pleomorphic xanthoastrocytoma (AUC 0.95).

• Preoperative multiparametric MRI from 1,844 patients with gliomas or glioneuronal/neuronal tumors was used to train and externally validate a MobileNetV2-based three-level classifier.
• Sequential forward feature selection identified optimal sequence combinations for each classification step, including T1 postcontrast, T2 FLAIR, T2, and ADC.
• Integrated per-category AUCs were 0.92, 0.81, 0.87, 0.82, 0.94, and 0.69, with accuracies of 0.88, 0.88, 0.89, 0.81, 0.91, and 0.95.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.