AI-augmented intraoperative decision-making workflows in diffuse midline glioma biopsy using cryosection pathology

Abstract Cryosection pathology is essential for intraoperative diagnosis of diffuse midline gliomas, yet it often leads to diagnostic errors and may prompt unnecessary re-biopsies before completion of the formal molecular assessment. In this study, we propose an AI-augmented framework, CryoAID, for rapid molecular outcome prediction during surgery for patients with diffuse midline glioma. CryoAID integrates a generative model to correct cryosection artefacts and a pathology foundation model to predict molecular statuses directly from cryosection images. We validate CryoAID across multiple cohorts to predict tumoural molecular statuses in the internal ( n = 326), external multi-centre ( n = 52), and consecutive ( n = 68) datasets. In particular, CryoAID accurately predicts major molecular statuses (e.g., ATRX, H3K27M, and TP53) using cryosection images that were previously deemed disqualified for molecular examinations. Beyond tumour cells, CryoAID reveals highly differential clinical features, including glial cell proliferation, abundant cytoplasm, and localised endothelial proliferation. In the retrospective analyses, CryoAID reduces re-biopsy rates by 26.4% and 26.6% in the internal and consecutive datasets, respectively. Our findings demonstrate that the AI-augmented pathology workflow can extract diagnostic value from specimens previously considered non-viable by traditional histopathology. This approach represents a shift towards real-time molecular pathology, potentially reducing re-biopsies and improving diagnostic precision for patients with diffuse midline glioma.