MDS-ATCN: Multiscale Depth-wise Separable Adaptive Temporal Convolutional Network for Predicting Brain Stroke using MRI images

Stroke is the leading cause of adult impairments, producing health, social, and economic issues for senior persons. Timely stroke risk prediction improves patient recovery by prompting medical treatments. Magnetic resonance imaging (MRI) of the brain provides important structural and temporal information, making it an important tool for evaluating strokes. The ability of Logistic regression (LR), Support vector machine (SVM), and LightGBM (LGBM), to understand complex spatial-temporal dependencies in MRI data remains restricted since they depend on both manually crafted features and predefined parameters. The proposed framework, named Multiscale depth-wise separable adaptive temporal convolutional network (MDS-ATCN), builds a deep learning model that addresses both issues by designing an effective stroke prediction method. The model combines diverse feature extraction techniques, including depth-wise separable convolutions with adaptive filters using attention mechanisms, to analyze temporal patterns while emphasizing significant brain structures. This study uses the ATLAS MRI dataset to evaluate the effectiveness of the proposed MDS-ATCN. This model achieved the highest accuracy of 94.2%, has a higher F1 score of 93.5%, and has a ROC-AUC higher than 95.1% compared to other machine learning models. It also proves that MDS-ATCN can produce accurate, large-scale predictions. Thus, it is crucial for clinical diagnosis by healthcare providers.