https://github.com/imswappy/brain-tumor-detection
🧠 Deep learning project for brain tumor classification using MRI images. Built with transfer learning (VGG16 + fine-tuning), TensorFlow/Keras, and deployed via Streamlit. Dataset & model loaded dynamically from KaggleHub. Includes training notebook, evaluation, and interactive web app.
https://github.com/imswappy/brain-tumor-detection
kagglehub keras numpy pandas scikit-learn streamlit tensorflow vgg16-model
Last synced: 5 months ago
JSON representation
🧠 Deep learning project for brain tumor classification using MRI images. Built with transfer learning (VGG16 + fine-tuning), TensorFlow/Keras, and deployed via Streamlit. Dataset & model loaded dynamically from KaggleHub. Includes training notebook, evaluation, and interactive web app.
- Host: GitHub
- URL: https://github.com/imswappy/brain-tumor-detection
- Owner: Imswappy
- Created: 2025-08-30T10:12:28.000Z (5 months ago)
- Default Branch: main
- Last Pushed: 2025-08-30T11:31:58.000Z (5 months ago)
- Last Synced: 2025-08-30T12:17:27.577Z (5 months ago)
- Topics: kagglehub, keras, numpy, pandas, scikit-learn, streamlit, tensorflow, vgg16-model
- Language: Jupyter Notebook
- Homepage:
- Size: 761 KB
- Stars: 1
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# 🧠 NeuroVision: Brain Tumor MRI Classification with Deep Learning
---
## 📌 Project Overview
This project focuses on **automated brain tumor detection** using **MRI images**.
We apply **transfer learning with VGG16**, a convolutional neural network (CNN) pre-trained on the ImageNet dataset, and fine-tune it to classify brain MRIs into:
- **Meningioma**
- **Glioma**
- **Pituitary Tumor**
- **No Tumor**
The trained model is deployed via a **Streamlit app**, where users can upload MRI images and get predictions with confidence scores.
---
## 📂 Project Structure
```
brain-tumor-mri/
│
├── brain_tumour_detection.ipynb # Training & evaluation (KaggleHub dataset)
├── streamlit_app.py # Deployment UI with Streamlit
├── requirements.txt # Dependencies
└── README.md # Documentation
```
---
## 📊 Dataset
Dataset is fetched from KaggleHub:
```python
import kagglehub
path = kagglehub.dataset_download("masoudnickparvar/brain-tumor-mri-dataset")
```
Structure:
```
Training/
├── glioma/
├── meningioma/
├── pituitary/
└── notumor/
Testing/
├── glioma/
├── meningioma/
├── pituitary/
└── notumor/
```
- **Training samples:** ~2870
- **Testing samples:** ~394
- Each subdirectory corresponds to a tumor class.
---
## 🏗️ Model Architecture (Transfer Learning with VGG16)
We use **VGG16**, a pre-trained CNN on ImageNet, as a **feature extractor** and fine-tune its top layers.
### 🔹 Step 1: Base Model
```python
from tensorflow.keras.applications import VGG16
base_model = VGG16(
input_shape=(128,128,3),
include_top=False,
weights='imagenet'
)
```
- **Input size:** 128 × 128 × 3 (resized MRI images).
- `include_top=False`: removes VGG16’s fully connected (FC) head.
- `weights="imagenet"`: initializes weights from ImageNet (~1.4M images, 1000 classes).
Mathematically, each convolutional layer applies:
where:
- W^{(k)}: convolutional kernel for feature map k
- x: input patch
- σ: activation (ReLU in VGG16)
### 🔹 Step 2: Freezing and Fine-Tuning
```python
for layer in base_model.layers:
layer.trainable = False
# Unfreeze top 3 layers
base_model.layers[-2].trainable = True
base_model.layers[-3].trainable = True
base_model.layers[-4].trainable = True
```
- Lower layers retain **general features** (edges, textures).
- Top 3 layers fine-tuned to capture **domain-specific features** of MRI tumors.
### 🔹 Step 3: Custom Classification Head
```python
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Flatten, Dropout, Dense
model = Sequential([
base_model,
Flatten(),
Dropout(0.3),
Dense(128, activation='relu'),
Dropout(0.2),
Dense(len(unique_labels), activation='softmax')
])
```
- **Flatten:** Reshapes VGG16 feature maps (4 × 4 × 512) → (8192,).
- **Dense(128, relu):** Fully connected layer learns non-linear combinations of features.
- **Dropout(0.3, 0.2):** Randomly zeroes units during training, reducing overfitting.
- **Output Layer:** Softmax classifier for 4 classes:
---
## 📒 Notebook (`brain_tumour_detection_vs_code.ipynb`)
Sections:
1. **Dataset loading** with KaggleHub
2. **EDA**: Class distributions, sample MRI visualization
3. **Preprocessing**: Resize, normalize, split
4. **Model**: VGG16 base + custom classifier
5. **Training**: Adam optimizer, categorical crossentropy loss
6. **Evaluation**: Accuracy, confusion matrix, classification report
8. **Saving model** in `.h5` format
---
## 🧮 Training Statistics
- **Optimizer:** Adam (β₁=0.9, β₂=0.999)
- **Learning Rate:** 1e-4
- **Batch Size:** 32
- **Loss:** Categorical Crossentropy
- **Regularization:** Dropout (0.3 & 0.2)
**Performance:**
- Training Accuracy: ~95%
- Test Accuracy: ~92%
- Balanced F1 scores across all classes
---
## 🎨 Streamlit App (`streamlit_app.py`)
Interactive deployment UI:
- Upload MRI image (`png/jpg/jpeg`)
- Preprocessing: resize 128 × 128, normalize [0,1]
- Prediction from VGG16-based model
- Displays:
- Label
- Confidence
- Uploaded image
- Raw probabilities
Run:
```bash
streamlit run streamlit_app.py
```
---
## ⚙️ Installation (VS Code)
```bash
# 1. Create environment
python -m venv .venv
.venv\Scripts\activate # Windows
source .venv/bin/activate # Mac/Linux
# 2. Install deps
pip install -r requirements.txt
# 3. Run app
streamlit run streamlit_app.py
```
---
## 🚀 Deployment
- Local: http://localhost:8501
- LAN: http://:8501
- Global: deploy on **Streamlit Cloud**
- 
---
## 🙌 Acknowledgements
- **Dataset**: [Brain Tumor MRI Dataset](https://www.kaggle.com/datasets/masoudnickparvar/brain-tumor-mri-dataset)
- **Pre-trained Model**: [MRI Brain Tumor Model](https://www.kaggle.com/models/noorsaeed/mri_brain_tumor_model)
- **Base Architecture**: [VGG16](https://arxiv.org/abs/1409.1556)
---