abdoalsir/african-neuroimaging-ai

Evaluating deep learning brain tumor classification performance on African neuroimaging data with Grad-CAM explainability and offline mobile deployment # African Neuroimaging AI: Evaluating Brain Tumor Classification Under Domain Shift **Evaluating deep learning brain tumor classification performance on African neuroimaging data, with Grad-CAM explainability analysis and offline mobile deployment.** ## The Question When a CNN is trained on high-field Western brain tumor MRI data, what happens when it's shown low-field African MRI data? Does accuracy degrade — and if it does, do the model's explanations (Grad-CAM) still make anatomical sense, or do they start pointing at scanner artifacts instead of pathology? No published study has measured this for brain tumor classification specifically. BraTS-Africa (Adewole et al., 2023) exists but covers segmentation only. No published study has evaluated Grad-CAM reliability on African MRI. No offline-capable mobile deployment tool exists for this use case in African clinical settings. This project addresses all three. ## Why This Project Exists I'm a recent medical graduate (MBBS, Sudan University of Science and Technology, 2026) with clinical training in neurology and neuroscience, working through Sudan's active conflict and its collapsed neuroimaging infrastructure (fewer than 20% of pre-war health facilities are still operational). Most brain tumor classification research is built and validated on Western datasets by computer science researchers without clinical grounding. This project is built by someone who has to answer the question "what does a wrong classification cost a patient" as a starting assumption, not an afterthought. ## Research Questions **Primary:** When a CNN trained on high-field Western brain tumor MRI data is evaluated on low-field African MRI data, what is the magnitude of the classification performance gap, and do Grad-CAM explainability maps remain clinically meaningful under this domain shift? **Secondary:** - Which tumor classes show the greatest performance degradation under domain shift? - Do Grad-CAM maps highlight pathology or scanner …