Classification of histology and molecular subtypes of brain gliomas and glioneuronal and neuronal tumors using a deep learning approach.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China.
- School of Medical Technology, Beijing Institute of Technology, Beijing, China.
- Philips Healthcare, Shanghai, China.
- Department of Radiology, West China Hospital of Sichuan University, Chengdu, China.
- Department of Radiology, The First Affiliated Hospital of Anhui Medical University, Hefei, China.
- School of Medical Technology, Beijing Institute of Technology, Beijing, China. [email protected].
- School of Medical Technology, Beijing Institute of Technology, Beijing, China. [email protected].
- Department of Radiology, Beijing Tiantan Hospital, Capital Medical University, Beijing, China. [email protected].
Abstract
To develop a deep learning classification system for integrated histology and molecular subtyping of gliomas and glioneuronal/neuronal tumors (GNTs). Preoperative multi-parametric MRI data from 1,844 patients with gliomas or GNTs, encompassing T1WI, post-contrast T1WI (T1(+C)), T2Flair, T2WI, and DWI, were analyzed. Subjects were classified into six categories: (1) oligodendroglioma, IDH-mutant, and 1p/19q-codeleted; (2) astrocytoma, IDH-mutant; (3) glioblastoma, IDH-wildtype; (4) pilocytic astrocytoma (PA); (5) pleomorphic xanthoastrocytoma (PXA); and (6) GNTs. A three-level classification system (MRI-G/GNTs) using MobileNetV2 was trained and validated in internal and external datasets. Sequential forward feature selection (SFFS) was utilized to optimize multi-modal network combinations. Furthermore, the MRI-G/GNTs outputs were integrated into structured virtual pathology reports. The average Dice for the brain tumor segmentation model was 0.92. Using SFFS, we identified optimal MRI sequence combinations per classification step. The external tests AUCs reached 0.92 for distinguishing adult-type diffuse gliomas from circumscribed astrocytic gliomas/GNTs (T1(+C), Flair, and T2WI combination), 0.92 for distinguishing IDH status for adult-type diffuse glioma (T1(+C)/Flair/ADC), 0.83 for distinguishing 1p/19q status for IDH-mutant diffuse glioma (ADC/Flair), 0.84 for circumscribed astrocytic gliomas vs GNTs (ADC/T1(+C)), 0.95 for PA vs PXA (ADC/T1(+C)/T2WI). Based on the automated prediction pipeline, the integrated classification performance of the MRI-G/GNTs system achieved AUCs of 0.92, 0.81, 0.87, 0.82, 0.94, and 0.69, with accuracy of 0.88, 0.88, 0.89, 0.81, 0.91, and 0.95 across the above six tumor categories. The MRI-G/GNTs system demonstrates strong capability in distinguishing both histological and molecular subtypes of brain gliomas and GNTs. Question How about the deep learning model based on MRI for integrated histology and molecular classification of brain GNTs, and can it be improved? Findings The MRI-G/GNTs system achieved AUCs of 0.69-0.92 with accuracy of 0.81-0.95 for distinguishing six tumor categories of GNTs. Clinical relevance The three-level deep learning classification system (MRI-G/GNTs) based on optimized multimodal MRI integration firstly achieves comprehensive histology and molecular subtype classification of brain GNTs.