Related Experiment Video
Updated: Apr 28, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
48.8K
MRI-based brain stroke classification using a hybrid vision transformer-BiLSTM architecture
Reeta Samuel1, Thanapal Pandi1
1School of Computer Science Engineering and Information Systems, Vellore Institute of Technology, Vellore, Tamil Nadu, India.
Frontiers in Neurology
|April 27, 2026
Summary
A new hybrid deep learning model combining Vision Transformer (ViT) and Bidirectional Long Short-Term Memory (BiLSTM) significantly improves stroke detection accuracy from MRI scans. This advanced AI approach enhances early stroke diagnosis and clinical decision support.
Area of Science:
- Artificial Intelligence in Medical Imaging
- Deep Learning for Neurological Disorders
- Biomedical Signal Processing
Background:
- Stroke is a leading cause of long-term disability, necessitating accurate and early detection methods.
- Hemorrhagic and ischemic strokes present diagnostic challenges due to variations in appearance and location.
- Magnetic Resonance Imaging (MRI), especially Diffusion-Weighted Imaging (DWI), is crucial for early stroke detection due to its sensitivity to fluid balance changes.
Purpose of the Study:
- To develop and evaluate a hybrid deep learning model for automated stroke classification using MRI data.
- To compare the performance of the hybrid model against established deep learning architectures.
- To assess the model's potential for clinical decision support in early stroke diagnosis.
Main Methods:
- A hybrid model integrating Vision Transformer (ViT) for spatial feature extraction and Bidirectional Long Short-Term Memory (BiLSTM) for temporal dependency modeling was developed.
- ViT utilized patch-based self-attention (16x16 patches, 256-D projections, 4 encoder layers, 8 attention heads) to capture global representations.
- The hybrid architecture was compared against baseline Convolutional Neural Network (CNN), VGG16, ResNet50, standalone ViT, and standalone BiLSTM models.
Main Results:
- The hybrid ViT-BiLSTM model achieved superior performance with 97.35% precision, 93.04% recall, 95.21% accuracy, 95.15% F1-score, and 99.36% ROC-AUC.
- The standalone ViT model (91.3% accuracy) outperformed CNN-based methods, highlighting the effectiveness of transformer architectures.
- Five-fold cross-validation demonstrated the hybrid model's stable performance, with an average accuracy of 96.61% (SD=0.78).
Conclusions:
- The ViT-BiLSTM model effectively combines global spatial context from ViT with bidirectional temporal features from BiLSTM for enhanced stroke classification from MRI.
- This hybrid approach shows significant promise for integration into clinical decision support systems for early stroke diagnosis.
- Future research directions include exploring federated learning for privacy preservation and assessing model generalizability across multi-institutional datasets.
