---

  What is NeuroScan?

NeuroScan is a full-stack medical imaging application that classifies brain MRI scans into four categories using a fine-tuned MobileNetV2 model. Upload an MRI, and the app instantly tells you the tumor type, shows a Grad-CAM heatmap highlighting the affected region, and validates the image is actually an MRI using OpenAI CLIP — all in under 2 seconds.

---

  Features

4-Class Classification Identifies Glioma, Meningioma, Pituitary tumors, or a healthy scan with 96.8% accuracy.	Grad-CAM Heatmaps Generates visual saliency maps that highlight exactly which brain regions influenced the model's prediction.
CLIP MRI Validation Uses OpenAI CLIP to reject non-MRI images before they reach the classifier — no garbage in, no garbage out.	Confidence Scores Returns a full probability breakdown across all four classes for every prediction.
Random Sample Testing One-click testing using real MRI samples bundled with the application — no upload required.	Dark Mode UI Sleek, responsive interface with smooth animations — designed for clarity and ease of use.

---

  Model Architecture

Base Model	MobileNetV2 — ImageNet pre-trained
Input	224 × 224 RGB
Backbone	Depthwise separable convolutions
Head	GAP → Dropout(0.5) → Dense(4) → Softmax
Accuracy	96.8% on held-out test set
Inference	< 2s on CPU

---

  API Reference

Endpoint	Method	Description
/	GET	API info & version
/health	GET	Health check
/stats	GET	Model & system stats
/predict	POST	Classify an uploaded MRI
/heatmap	POST	Generate Grad-CAM heatmap
/random	GET	Test with a random sample MRI

---

  Tech Stack

	Technologies
Backend	Python · Flask · TensorFlow/Keras · OpenCV · CLIP
Frontend	HTML · CSS · Vanilla JavaScript
Hosting	Hugging Face Spaces (backend) · Vercel (frontend)

---

  Dataset

The Brain Tumor Classification MRI dataset from Kaggle — 3,264 labeled MRI scans across four classes.

Class	Count	Description
Glioma	~826	Primary tumors from glial cells
Meningioma	~822	Tumors from the meninges
Pituitary	~827	Tumors of the pituitary gland
No Tumor	~395	Healthy brain scans

---

  Run Locally

# 1. Clone
git clone https://github.com/yashnaiduu/NeuroScan-Brain-Tumor-Classification.git
cd NeuroScan-Brain-Tumor-Classification

# 2. Environment
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# 3. Start backend
python server1.py # → http://localhost:5050

# 4. Start frontend (new terminal)
cd client && python3 -m http.server 8000 # → http://localhost:8000

Tip: Set PORT=7860 when deploying to Hugging Face Spaces.

---

  Deployment

Platform	Role	Key Files
Hugging Face Spaces	Backend API	Dockerfile, entrypoint.sh
Vercel	Frontend	vercel.json

Detailed steps in DEPLOYMENT.md.

---

  License

MIT — see LICENSE.

---

Yash Naidu

_{Built with TensorFlow  ·  Deployed on Hugging Face & Vercel}