3D Resnet-18 Based Deep Convolutional Network for Automated Stroke Detection in Volumetric CT Scans

Stroke detection using CT imaging presents significant challenges due to the subtle and heterogeneous nature of stroke lesions, variations in imaging protocols, and limited availability of large annotated datasets. Traditional methods often rely on $2 D$ image analysis or handcrafted features, which inadequately capture the volumetric and spatial context critical for accurate stroke identification. Additionally, class imbalance between stroke and non-stroke cases frequently leads to biased models with poor sensitivity, compromising clinical reliability. This work proposes a 3D ResNet-18 deep learning framework tailored for volumetric CT brain scans to overcome these challenges. By leveraging three-dimensional convolutions, the model effectively exploits spatial correlations across consecutive slices, enabling robust identification of both ischemic and hemorrhagic strokes. The use of residual connections facilitates stable training and deeper feature extraction, addressing vanishing gradient problems commonly encountered in volumetric models. To enhance model generalization, extensive preprocessing including voxel resampling, intensity normalization, and data augmentation are applied, mitigating variability across datasets. Class imbalance is explicitly handled through weighted loss functions, improving sensitivity critical for clinical diagnosis. Key features include the end-to-end volumetric analysis that preserves spatial context, interpretability via Grad-CAM visualization to highlight clinically relevant brain regions, and rigorous evaluation through stratified $\mathbf{k}$-fold cross-validation ensuring unbiased performance metrics. The proposed methodology demonstrates high accuracy and sensitivity in detecting subtle stroke lesions, positioning it as a viable clinical decision support tool in emergency stroke diagnosis workflows.