Brain Tumor Detection & Classification

An end-to-end deep learning system for automated brain tumor detection and classification using MRI scans. Built using Transfer Learning (ResNet50) with Grad-CAM explainability and deployed via a Streamlit web application.

---

Project Overview

Brain tumor diagnosis requires expert radiologists to manually examine MRI scans — a process that is time-consuming, expensive, and prone to human error. This project automates tumor detection and classification using deep learning, providing instant predictions with visual explanations to assist clinical decision-making.

Tumor Classes:

• 🔴 Glioma
• 🟡 Meningioma
• 🟢 Pituitary Tumor
• ⚪ No Tumor

---

Team

Name	Role
Nishit Patel	Model Building & Training
Pranav Adhikari	Model Evaluation & Metrics
Unique Bhakta Shrestha	EDA & Visualizations
Pragun Lal Shrestha	Data Preprocessing & Augmentation

Grad-CAM explainability and Streamlit deployment are shared across the team.

---

Project Structure

brain_tumor_detection/
├── dataset/                  # MRI images (not pushed to GitHub)
│   ├── Train/
│   │   ├── glioma/
│   │   ├── meningioma/
│   │   ├── notumor/
│   │   └── pituitary/
│   └── Test/
│       ├── glioma/
│       ├── meningioma/
│       ├── notumor/
│       └── pituitary/
├── notebooks/
│   ├── 01_eda.ipynb               # Exploratory Data Analysis
│   ├── 02_preprocessing.ipynb     # Preprocessing & Augmentation
│   ├── 03_model_training.ipynb    # Model Building & Training 
│   └── 04_evaluation.ipynb        # Evaluation & Grad-CAM
├── models/
│   ├── final_model.keras          # Best trained model 
│   └── class_indices.json         # Class label mapping
├── app/
│   └── app.py                     # Streamlit web application
├── utils/
│   └── gradcam.py                 # Grad-CAM implementation
├── .gitignore
├── requirements.txt
└── README.md

---

Dataset

Source: Mendeley Data — Brain Tumor MRI Dataset
Link: https://data.mendeley.com/datasets/zwr4ntf94j/5
Total Images: 12,064 MRI scans
Modality: T1-weighted contrast-enhanced MRI

Split	Glioma	Meningioma	No Tumor	Pituitary	Total
Train	3,018	2,183	1,945	2,504	9,650
Test	755	626	487	546	2,414

⚠️ Dataset is not included in this repository due to size. Download from the link above and place in the dataset/ folder.

---

Tech Stack

Category	Tools
Language	Python 3.10
Deep Learning	TensorFlow 2.10.0, Keras
Model	ResNet50
Explainability	Grad-CAM
Data Processing	NumPy, Pandas, OpenCV, Pillow
Visualization	Matplotlib, Seaborn
Evaluation	Scikit-learn
Deployment	Streamlit

---

Project Workflow

Data Collection → EDA → Preprocessing & Augmentation → Model Building
      → Training (2-Phase) → Evaluation → Grad-CAM → Streamlit Deployment

1. EDA

• Class distribution analysis
• Sample image visualization per class
• Image dimension analysis
• Pixel intensity distribution

2. Preprocessing & Augmentation

• Resize to 224×224 pixels
• Normalize pixel values (0–1)
• Augmentation: flip, rotate, zoom, brightness shift
• Class weights to handle imbalance
• 80/20 train/validation split

3. Model — ResNet50 + Transfer Learning

• Phase 1: Frozen base, train custom head (lr=1e-3, 15 epochs)
• Phase 2: Fine-tune top 20 layers (lr=1e-5, 20 epochs)
• Callbacks: EarlyStopping, ModelCheckpoint, ReduceLROnPlateau

4. Evaluation

• Accuracy, Precision, Recall, F1-Score
• Confusion Matrix
• Classification Report per class
• Emphasis on Recall to minimize false negatives

5. Grad-CAM Explainability

• Heatmap overlay on MRI showing regions influencing prediction
• Builds clinical interpretability and trust

6. Streamlit Deployment

• Upload MRI image
• Get instant prediction with confidence score
• View Grad-CAM heatmap

---

Setup & Run

Prerequisites

python 3.10

Installation

git clone https://github.com/nish-debug15/brain_tumor_detection.git
cd brain_tumor_detection
python -m venv brain_tumor_env
brain_tumor_env\Scripts\activate
pip install -r requirements.txt

Run Notebooks

Open in VS Code with Jupyter extension, run in order:

1. 01_eda.ipynb
2. 02_preprocessing.ipynb
3. 03_model_training.ipynb
4. 04_evaluation.ipynb

Run Streamlit App

cd app
streamlit run app.py

---

Results

Metric	Score
Training Accuracy	74.67%
Validation Accuracy	76.82%
Test Accuracy	74.19%
Precision (Weighted)	73.55%
Recall (Weighted)	74.19%
F1 Score (Weighted)	72.84%

Per Class Performance

Class	Precision	Recall	F1 Score
Glioma	81%	61%	69%
Meningioma	55%	45%	50%
No Tumor	76%	98%	85%
Pituitary	80%	97%	87%

Model: ResNet50 with Transfer Learning. Best performance on No Tumor (98% recall) and Pituitary (97% recall) classes. Meningioma remains the most challenging class due to visual similarity with other tumor types.

---

Key Design Decisions

Why ResNet50? ResNet50 achieves strong performance on medical imaging tasks through residual/skip connections that solve the vanishing gradient problem, enabling effective training of deep 50-layer networks. We initially tried EfficientNetB0 but it plateaued at 33% accuracy on MRI data — ResNet50 reached 68% by Epoch 2, making it the clear choice for this dataset.

Why Transfer Learning?
Our dataset of 12,064 images is insufficient to train a CNN from scratch. ResNet50 pre-trained on ImageNet (1.2M images) provides rich feature representations that transfer well to medical imaging.

Why Grad-CAM?
Medical AI without explainability cannot be trusted clinically. Grad-CAM makes the model's decision process transparent by highlighting the exact MRI regions that influenced the prediction.

Why emphasize Recall?
A false negative (missing a real tumor) is far more dangerous than a false positive. We optimize recall to minimize the risk of undetected tumors.

---

License

This project is for academic purposes — 4th Semester AIML Project.