🧠 NeuroScan AI — Brain Tumor Detection

An AI-powered MRI brain scan classifier built with EfficientNet-B0, achieving 99.39% test accuracy across four diagnostic categories. Includes Grad-CAM explainability and an interactive Streamlit dashboard.

---

📌 Project Overview

NeuroScan AI is a deep learning system that classifies brain MRI scans into four categories:

• Glioma — a type of tumor originating in the brain or spinal cord
• Meningioma — a tumor arising from the meninges
• Pituitary — a tumor located at the base of the brain
• No Tumor — a healthy scan with no detected abnormality

The model uses transfer learning on a pretrained EfficientNet-B0 backbone, fine-tuned with heavy data augmentation on a labeled MRI dataset. A Grad-CAM heatmap overlay provides visual explainability, highlighting the exact regions of the MRI that influenced the model's prediction.

---

🗂️ Project Structure

NeuroScan-AI/
│
├── data/
│   ├── Training/          # Training images (organized by class)
│   └── Testing/           # Test images (organized by class)
│
├── models/
│   ├── brain_tumor_efficientnet.pth            # Base trained model
│   ├── brain_tumor_efficientnet_finetuned.pth  # Fine-tuned model
│   └── brain_tumor_efficientnet_augmented.pth  # Final augmented model (used in app)
│
├── train.py               # Initial training from pretrained EfficientNet-B0
├── finetune.py            # Fine-tuning with augmentation on trained model
├── augmentedFinetune.py   # Advanced fine-tuning with clean val split strategy
├── evaluate.py            # Full evaluation with confusion matrix & report
├── predict.py             # Single-image prediction with Grad-CAM visualization
└── app.py                 # Streamlit web application

---

⚙️ Model Training Pipeline

The model was developed in three progressive stages:

Stage 1 — train.py

• Loads pretrained EfficientNet-B0 from timm (ImageNet weights)
• Replaces the classification head with a 4-class output layer
• Trains for 5 epochs with AdamW optimizer (LR: 0.001)
• 80/20 train/validation split
• Saves weights to models/brain_tumor_efficientnet.pth

Stage 2 — finetune.py

• Loads Stage 1 weights and continues training
• Adds RandomHorizontalFlip and RandomRotation augmentation
• Runs for 10 more epochs with a lower LR (0.0001)
• Adds weight_decay=1e-4 to reduce overfitting
• Saves to models/brain_tumor_efficientnet_finetuned.pth

Stage 3 — augmentedFinetune.py (Final Model)

• Loads Stage 1 weights and applies heavy augmentation to training data:
  • RandomHorizontalFlip, RandomRotation(15), RandomAffine, ColorJitter
• Validation is kept clean (no augmentation) for honest evaluation
• Uses a deterministic index-based train/val split via numpy
• Saves to models/brain_tumor_efficientnet_augmented.pth

---

📊 Evaluation — evaluate.py

Runs the final augmented model against the held-out Testing directory.

Outputs:

• Overall test accuracy
• Per-class classification_report (precision, recall, F1)
• Confusion matrix heatmap

Result: ✅ 99.39% Test Accuracy

---

🔍 Grad-CAM Explainability

Both predict.py and app.py use Gradient-weighted Class Activation Mapping (Grad-CAM) to generate heatmaps that visually explain model predictions.

• Target layer: model.conv_head (final convolutional layer of EfficientNet)
• Highlights regions in the MRI that most strongly influenced the classification
• Rendered as a color overlay on the original scan

---

🖥️ Streamlit App — app.py

An interactive web dashboard for non-technical users.

Features:

• Upload any MRI scan (JPG/PNG)
• Instant classification with confidence score
• Per-class confidence breakdown with visual progress bars
• Grad-CAM heatmap overlay with explainability section
• Dark-themed, responsive UI

Run the app:

streamlit run app.py

---

🛠️ Installation

Requirements:

pip install torch torchvision timm streamlit streamlit-lottie \
            pytorch-grad-cam scikit-learn matplotlib pillow requests

Dataset:
Download the Brain Tumor MRI Dataset from Kaggle and place it in ./data/ with Training/ and Testing/ subdirectories.

Training from scratch:

python train.py
python augmentedFinetune.py
python evaluate.py

---

🧪 Quick Prediction Test

python predict.py

Randomly selects an image from the test set, runs inference, and displays the original scan alongside the Grad-CAM heatmap — color-coded green (correct) or red (incorrect).

---

📈 Results Summary

Model Version	Val Accuracy
Base (train.py)	~96–97%
Fine-tuned	~98%
Augmented (final)	99.39%

---

🏗️ Tech Stack

Component	Technology
Model	EfficientNet-B0 (timm)
Framework	PyTorch
Explainability	Grad-CAM (pytorch-grad-cam)
UI	Streamlit
Data	torchvision.datasets.ImageFolder
Hardware	Apple MPS / CUDA / CPU

---

⚠️ Disclaimer: NeuroScan AI is a research and educational project. It is not intended for clinical use or to replace professional medical diagnosis.