Two-stage classification of glioma subtypes using a hybrid CNN-ViT model

Gliomas are among the most serious brain tumor types due to their aggressive behavior. Recent advances in automated methods for early brain tumor detection that analyze whole-slide images of brain tissue enable rapid and precise diagnosis. Therefore, we propose a hybrid approach combining a convolutional neural network and a vision transformer for deep analysis of whole-slide patches. This hybrid model employs a two-stage classification strategy to address highly imbalanced class distributions. The first stage uses a binary-label classifier to label patches as either normal or abnormal. The second stage employs a multi-label classifier to classify abnormal patches into their corresponding glioma subtypes. We evaluated the two-stage hybrid model on an imbalanced test set, achieving the highest balanced accuracy (95.31%), macro-averaged precision (96.08%), and macro-averaged F1-score (95.51%). The two-stage classification approach addressed class imbalance. Furthermore, the hybrid model, which combines a convolutional neural network and a vision transformer, captured deeper features, achieving superior performance and outperforming state-of-the-art methods. المزيد