EffNet-ViT: An Interpretable Hybrid EfficientNet-Transformer Architecture for Ischemic Stroke Classification on CT Scans

Abstract

The detection and early, correct classification of brain stroke on non-contrast CT (NCCT) is essential to timely intervention, but remains challenging because of the insidious and diffuse characteristics of early ischemic lesions. While standard Convolutional Neural Networks (CNNs) have proven effective, this study proposes EffNet-ViT, a new interpretable hybrid architecture, to classify strokes into three classes (Normal, Ischemia, Bleeding). The model combines a pre-trained EfficientNet-B3 backbone, which learns efficient local features, with a Transformer encoder head that captures long-distance global dependencies in the images. EffNet-ViT was trained and tested on 6,653 NCCT scans using a 5-fold stratified cross-validation scheme, achieving an ensemble accuracy of 97.94%. On an external held-out test set the model confirmed its robustness with 97.74% accuracy and a high F1-score in the critical Ischemia category. Grad-CAM explainable-AI visualizations provide clinical interpretability, showing that the model relies on clinically relevant radiological regions and thereby increasing trust and transparency. The proposed architecture is a robust, interpretable framework for automatic stroke detection and a valuable clinical decision-support tool.