Diffuse glioma classification with deep learning and explainability: addressing challenges in histopathology image analysis

Brain tumors are among the most lethal diseases, causing significant neurological damage and high mortality. Diffuse gliomas, a highly invasive subtype, challenge treatment due to widespread brain infiltration and therapy resistance. This paper proposes a deep learning framework for detecting diffuse glioma and its subtypes from H&E-stained WSI patches, with both patch-level and patient-level detection, along with visualization to enhance explainability. Automation of detection enhances speed and consistency, reducing the workload of pathologists significantly. The proposed framework consists of three stages: a diffuse glioma patch classifier, a diffuse glioma subtype classifier, and a patient-level classifier. Additionally, four different explainable AI techniques including, Grad-CAM, Grad-CAM++, Score-CAM, and Smooth Grad-CAM + + were employed to highlight the diffuse glioma-affected areas. This study utilizes the DeepHisto dataset consisting of tiles from H&E-stained WSIs of 28 adult-type diffuse glioma cases from the NCP, Luxembourg National Health Laboratory. The data were preprocessed using a novel technique and trained with classification models including DenseNet121, MobileNetV2, an ensemble of DenseNet121 and MobileNetV2, and Vision Transformer. The ensemble model showed robust performance in diffuse glioma patch classification, achieving a specificity of 96.14% and an accuracy of 91.33%. The diffuse glioma subtype classifier recorded a specificity of 92.26%. For patient-level identification, the diffuse glioma classifier achieved an F1 score of 93.53%, while the diffuse glioma subtype classifier reached an accuracy of 95.23% and a specificity of 97.61%. The proposed deep learning framework for the automated detection of diffuse glioma and its subtypes demonstrated promising results. This framework has the potential to serve as a reliable diagnostic tool, enabling efficient, accurate, and cost-effective analysis of diffuse gliomas.