[The Next Frontier of Medical AI in Brain Tumor Care:From Prediction to Trust].

Abstract

Artificial intelligence (AI) is rapidly reshaping the diagnosis and treatment of malignant brain tumors. Within the molecularly defined 2021 World Health Organization (WHO) framework, deep learning now supports several tasks: noninvasive, rapid prediction of molecular alterations from MRI and intraoperative optical imaging (radiogenomics); integrated interpretation of histology and genomics for prognosis; and accelerated treatment development through pathology foundation models, drug discovery, and large language model-based clinical trial matching and decision support. Across these advances, one recurring obstacle remains: distribution shift, or domain shift. Models whose performance depends on the provenance of their training data often perform less well at external institutions, and high discriminatory performance, as measured by the receiver operating characteristic area under the curve (ROC AUC), does not guarantee clinical reliability. We illustrate this issue through our study of IDH mutation prediction, which compared convolutional and transformer models with physicians using an international dataset and a Japanese cohort. External discrimination remained moderate, yet probability calibration deteriorated, whereas experienced clinicians maintained better calibration. We argue that the next frontier is not higher point-estimate accuracy but trustworthy AI: models that express well-calibrated uncertainty, abstain or defer in ambiguous cases, and request additional data within a human-in-the-loop workflow. Brain tumor care must move from prediction to trust.