NEUROSCAN AI Intelligent Brain Tumor Detection & Classification Using Deep Learning

A deep-learning classifier that reports a tumor type without saying how confident it is, which pixels led to that decision, or where the lesion sits is difficult to trust in a clinical workflow. We present NeuroScan AI, a brain-MRI decision-support system built to close that gap rather than to chase a single accuracy figure. The system classifies axial MRI slices into glioma, meningioma, pituitary tumor, or no-tumor using a dual-branch classifier that concatenates ResNet-50 and Swin-Tiny Transformer features ahead of a shared classification head; this simple-concatenation variant was selected over an attention-gated alternative through a four-way ablation. Every prediction carries a fifty-pass Monte Carlo Dropout uncertainty estimate, an Attention U-Net lesion segmentation, ten radiomics shape/texture features ranked by SHAP importance, and Grad-CAM-family explanations evaluated with Pointing-Game and Pixel-Flipping faithfulness metrics – all compiled into a deterministic, template-based clinical report that does not depend on a general-purpose language model. While auditing our own implementation for this paper, we found that the attention-gating branch in the ablated hybrid variant is computed but never applied to the feature vector it is meant to weight; we report this discrepancy openly rather than describing the architecture as fully attention-gated. The full pipeline is delivered inside a role-based, audited web application (FastAPI, React, SQLite) and was verified end to end by an automated test suite covering authentication, ingestion, reporting, and access control across three user roles, all passing. Following an evaluation protocol fixed before any test-set number was computed, the deployed classifier reached 86.18% test accuracy (macro precision 0.8674, macro recall 0.8531, macro F1 0.8576, ROC-AUC 0.9186), measurably below its training-set fit (accuracy 0.9147) and below the 99.8% validation AUC-ROC the application’s own landing page displays – a gap we surface and discuss rather than reconcile silently.