osmanyakubu-maker/HybridStrokeNet-Ghana-CT

HybridStrokeNet-Ghana-CT is an open-source explainable stroke classification solution using non-contrast brain CT scans. It employs EfficientNet-B3, Vision Transformer, Grad-CAM, and INT8 to provide a complete workflow that includes pre-processing, training, testing, and deployment. HybridStrokeNet is an interpretable hybrid CNN-Transformer architecture for automatic multi-class stroke classification in non-contrast brain computed tomography (CT) scans. The corresponding source code executes the approach presented in the paper, including image pre-processing, patient-wise data split, HybridStrokeNet training and evaluation, explainability by Grad-CAM and Attention Roll-out, external validation, and INT8 model export. This research used a large clinical dataset from Ghana, which consists of 12,843 CT slices of 634 patients and independent external validation of 4,215 CT slices of 210 patients. The goal is to contribute to reproducible research in the field of AI-aided stroke classification and help develop explainable AI models for resource-limited clinical practice.