MRI-based brain stroke classification using a hybrid vision transformer-BiLSTM architecture

Stroke is a prominent cause of long-term disability, impacting patients' socioeconomic status in daily life. Hemorrhagic and ischemic strokes differ in dimensions, forms, and locations, posing challenges for automated detection. Magnetic resonance imaging (MRI), particularly diffusion-weighted imaging (DWI), reveals changes in fluid balance, thereby enabling early detection. Hence, MRI scans are more accurate than computed tomography (CT) scans due to their increased sensitivity.

To categorize brain strokes, a hybrid model combining bidirectional long short-term memory (BiLSTM) with a vision transformer (ViT) was developed using an MRI dataset from a private source. ViT identifies qualities using MRI. The ViT captures global contextual and spatial representations using patch-based self-attention (16×16 patches, 256-dimensional projections, four transformer encoder layers with eight attention heads), whereas the BiLSTM network (128 and 64 units) models dependencies inside transformer-encoded features. A comparative study was conducted for the hybrid architecture with deep learning models, including a convolutional neural network (baseline, 85.5%), VGG16 (87.8%), ResNet50 (89.2%), ViT (91.3%), and BiLSTM (88.6%).

The hybrid ViT-Bi-LSTM model achieved a precision of 97.35%, recall of 93.04%, accuracy of 95.21%, F1-score of 95.15%, and ROC-AUC of 99.36%, outperforming other comparative approaches. The standalone ViT achieved an accuracy of 91.3%, exceeding the CNN-based methods. In 5-fold cross-validation, the hybrid ViT-BiLSTM model achieved an average accuracy of 96.61%, with a standard deviation of 0.78, indicating stable performance across folds. These findings validate the combination of bidirectional temporal modeling with transformer-based feature extraction.

By capturing the global spatial context through self-attention and bi-directional features via recurrent processing, ViT with Bi-LSTM networks expands stroke classification from MRI data. The ViT-Bi-LSTM model showed a promising approach for clinical decision support systems in early stroke diagnosis. Future research will use federated learning (FL) to protect privacy and assess model generalizability across multi-institutional MRI datasets.