2026AI & Deep Learning
NeuroDL
Full-Stack Brain Tumor Detection Platform
Overview
A full-stack brain tumor detection platform built as a final-year academic project. Classifies MRI scans into 4 classes with real-time inference, estimates prediction uncertainty, highlights the tumor region with Grad-CAM, and drafts a structured radiology-style report for a doctor to review.
Results
96.95% test accuracy on 1,311 held-out scans (trained on 7,000+ MRI images); AUC up to 0.9989 across the 4 classes.
Academic student project; not a clinical diagnostic tool. Performance figures are measured on public benchmark datasets.
How it works
- 01
Pipeline: MRI upload → 4-class classification (glioma, meningioma, pituitary, no tumor) → Grad-CAM heatmap and pseudo-segmentation overlay → LLM-drafted radiology report → persistent scan history.
- 02
ResNet50V2 + custom CNN ensemble, with real-time inference streamed to the client over Flask-SocketIO.
- 03
Monte Carlo Dropout uncertainty estimates shown alongside each prediction.
- 04
JWT authentication with separate patient and doctor roles, plus a doctor portal for reviewing scans.
- 05
Custom Grad-CAM pseudo-segmentation replaced an earlier U-Net, which didn't transfer because of a mismatch with its LGG training dataset.
- 06
DICOM ingestion via pydicom; radiology-style reports generated with the Groq API.