NeuroVisionAI: Bridging the Diagnostic and Explainability Gaps in Brain Hemorrhage Detection Using Artificial Intelligence

This work presents NeuroVisionAI, a deep learning-based framework for the detection and classification of intracranial hemorrhages from non-contrast CT (NCCT) scan images. The proposed approach integrates a convolutional neural network (CNN) architecture enhanced with a Convolutional Block Attention Module (CBAM) to improve feature representation and localization of hemorrhagic regions. A comprehensive preprocessing pipeline, including skull stripping, contrast enhancement (CLAHE), and intensity normalization, is employed to address variability in medical imaging data. The model is trained and evaluated on publicly available datasets such as CQ500 and RSNA Intracranial Hemorrhage datasets, along with curated clinical samples, using a stratified training and testing approach. To enhance interpretability, the framework incorporates Grad-CAM-based visual explanations, enabling transparent identification of regions influencing model predictions. Experimental results demonstrate strong performance across key evaluation metrics, including accuracy, sensitivity, and precision, indicating the model’s effectiveness in detecting multiple hemorrhage subtypes. The proposed system aims to support clinical decision-making by providing an interpretable and efficient AI-assisted diagnostic tool, contributing toward improved reliability and applicability of deep learning models in neuroimaging workflows.