What type of study is this?

This is a Quantitative Study study.

October 5, 2025

Ks02.6.a Computational Histopathology Enables Scalable Deep Learning-Driven CNS Tumor Classification for Real-Time Intraoperative Diagnostics

Key Points

The deep learning models achieved an accuracy of 85%, surpassing expert neuropathologists in diagnosing CNS tumors.
Utilizing convolutional neural networks, the system was trained on over 4,600 intraoperative whole slide images for robust classification.
Smear-based models distinguished various tumor types with 74% accuracy, addressing limitations when frozen sections are impractical.
This AI framework enhances intraoperative diagnostics in both resource-rich and limited environments, improving access to neuro-oncological care.

Abstract

Abstract BACKGROUND Accurate intraoperative diagnosis of central nervous system (CNS) tumors is critical for guiding neurosurgical decision-making. However, the global shortage of trained neuropathologists and limited access to molecular diagnostics—particularly in low-resource settings—pose significant challenges. Existing artificial intelligence (AI) solutions are predominantly limited to postoperative hematoxylin and eosin (H&E) slides or address a narrow spectrum of tumor entities, thereby restricting their clinical applicability and scalability for real-time intraoperative use. MATERIALS AND METHODS We developed deep learning-based classification models utilizing convolutional neural networks (CNNs) trained on a dataset of 4,699 intraoperative whole slide images (WSIs), including both frozen sections and smear preparations. All cases were annotated according to the 2021 WHO classification of CNS tumors. The models were trained to recognize 19 tumor classes, incorporating artifact detection modules and comprehensive data augmentation strategies to ensure robustness across different preparation methods, staining protocols, scanning systems, and image quality. RESULTS The frozen section classifier, based on the EfficientNetV2 architecture, achieved a diagnostic accuracy of 85% and outperformed expert neuropathologists on an independent test cohort. Smear-based models reliably distinguished astrocytic, oligodendroglial, and non-glial tumors, achieving 74% accuracy in classifying 10 distinct tumor types. These smear models offer a valuable diagnostic complement in cases where frozen section quality is compromised or unavailable. The complete AI pipeline has been integrated into a graphical user interface optimized for both specialist and non-specialist users in intraoperative settings. CONCLUSION This work presents a robust, scalable AI-based framework for intraoperative CNS tumor classification using widely accessible histopathological preparations. By delivering real-time diagnostic support without reliance on advanced molecular assays or specialist interpretation, this tool holds promise for support of intra-operative diagnostics in both high- and low-resource environments, advancing equitable access to high-quality neuro-oncological care.

Ask AI

Mark Helpful

Bookmark

Relay