Automated MRI Analysis Framework for Stroke Detection in Data-Limited Clinical Settings

This study presents an automated framework for T1-weighted MRI analysis in stroke diagnosis, addressing the challenges of data scarcity and class imbalance in clinical settings. We developed a controlled data augmentation strategy using synthetic templates generated from real images to create morphologically plausible representations consistent with pathological phenotypes. This approach expanded our dataset to 1910 cases while preserving anatomical structure and improving generalization. The framework utilizes a multimodal deep neural network that jointly processes MRI images and clinical data (age, NIHSS score, onset time, sex) through a balanced architecture designed for visual and tabular data integration. The model achieved strong performance: age prediction MAE of 4.03 years, NIHSS score MAE of 2.42, lesion classification accuracy of 98.0% with AUC 0.984 and F1-score 0.980 and overall test accuracy of 89.8%. Grad-CAM-based explainability components visualize regions influencing model decisions, providing interpretable insights for clinical integration. Correlation analyses confirm consistency between predictions and observed clinical values. This framework offers a practical solution for implementing deep learning-based stroke analysis in resource-constrained clinical environments, demonstrating how synthetic data augmentation can effectively address limited training data while maintaining clinical relevance and diagnostic accuracy.