Abstract
The latest World Health Organization (WHO) classification of central nervous system tumors (WHO2021/5th) has incorporated molecular information into the diagnosis of each brain tumor type including diffuse glioma. Therefore, an artificial intelligence (AI) framework for learning histological patterns and predicting important genetic events would be useful for future studies and applications. Using the concept of multiple-instance learning, we developed an AI framework named GLioma Image-level and Slide-level gene Predictor (GLISP) to predict nine genetic abnormalities in hematoxylin and eosin sections: IDH1/2, ATRX, TP53 mutations, TERT promoter mutations, CDKN2A/B homozygous deletion (CHD), EGFR amplification (EGFRamp), 7 gain/10 loss (7+/10-), 1p/19q co-deletion, and MGMT promoter methylation. GLISP consists of a pair of patch-level GLISP-P and patient-level GLISP-W models, each pair of which is for a genetic prediction task, providing flexibility in clinical utility. In this study, the Cancer Genome Atlas whole-slide images (WSIs) were used to train the model. A total of 108 WSIs from the Tokyo Women's Medical University were used as the external dataset. In cross-validation, GLISP yielded patch-level/case-level predictions with top performances in IDH1/2 and 1p/19q co-deletion with average areas under the curve (AUCs) of receiver operating characteristics of 0.75/0.79 and 0.73/0.80, respectively. In external validation, the patch-level/case-level AUCs of IDH1/2 and 1p/19q co-deletion detection were 0.76/0.83 and 0.78/0.88, respectively. The accuracy in diagnosing IDH-mutant astrocytoma, oligodendroglioma, and IDH-wild-type glioblastoma was 0.66, surpassing the human pathologist average of 0.62 (0.54-0.67). In conclusion, GLISP is a two-stage AI framework for histology-based prediction of genetic events in adult gliomas, which is helpful in providing essential information for WHO 2021 molecular diagnoses.