This repository contains the implementation of the term project for the course BIM453: Introduction to Machine Learning (2024-2025 Fall Semester). The goal of this project is to classify brain tumors from X-ray images into two categories: Healthy and Tumor using various deep learning techniques.
Brain tumor classification is critical in the medical field as timely and accurate diagnosis plays a significant role in treatment. This project explores the potential of deep learning in medical imaging by training and evaluating multiple models, including custom-designed convolutional neural networks (CNNs) and pre-trained architectures (ResNet50, MobileNetV2, and DenseNet121).
- Custom and Pre-trained Models: Implementation of two custom CNN architectures and three pre-trained models using transfer learning.
- Image Preprocessing: Techniques like histogram equalization, data augmentation, and normalization.
- Evaluation Metrics: Models are compared using metrics such as Accuracy, Precision, Recall, and F1-Score.
The dataset consists of X-ray images divided into two categories: Healthy and Tumor. The dataset was balanced to avoid bias in model predictions. Below are example images and the distribution of the dataset:
Before training the models, the dataset was preprocessed with the following techniques:
- Normalization
- Data Augmentation
- Histogram Equalization
- Basic CNN 1: A simple architecture with three convolutional layers, followed by max pooling and fully connected layers.
- Basic CNN 2: Enhanced version of CNN 1 with Batch Normalization and Dropout layers.
- ResNet50: Pre-trained on ImageNet, fine-tuned for binary classification.
- MobileNetV2: Lightweight and computationally efficient model, also fine-tuned for binary classification.
- DenseNet121: Utilizes dense connections for feature reuse, fine-tuned for the task.
- Frameworks: TensorFlow and Keras
- Optimizer: Adam with binary cross-entropy loss
- Platform: Google Colab with NVIDIA Tesla T4 GPUs
The performance of the models was evaluated using Accuracy, Precision, Recall, and F1-Score. The results are summarized below:
| Model | Accuracy | Precision | Recall | F1 Score |
|---|---|---|---|---|
| Basic CNN 1 | 95% | 95% | 95% | 95% |
| Basic CNN 2 | 87% | 87% | 86% | 87% |
| ResNet50 | 92% | 93% | 93% | 93% |
| MobileNetV2 | 91% | 93% | 91% | 92% |
| DenseNet121 | 92% | 93% | 91% | 92% |
├── data/
│ ├── raw/ # Original X-ray images
│ ├── processed/ # Preprocessed images
├── src/
│ ├── data_loading.py # Data preprocessing functions
│ ├── model.py # Model architectures
│ ├── visualization.py # Visualization utilities
│ ├── utils.py # Helper functions
├── notebooks/
│ ├── exploratory_data_analysis.ipynb
│ ├── modelling.ipynb
│ ├── inference.ipynb
├── experiments/ # Experimental results
├── models/ # Trained model weights
├── results/ # Visualizations of metrics
├── requirements.txt # Dependencies
Install the required dependencies using the following command:
pip install -r requirements.txt- Clone the repository:
git clone https://github.com/username/brain-tumor-classification.git
- Navigate to the directory:
cd brain-tumor-classification - Run the preprocessing and training scripts in the
notebooks/directory.
The study demonstrates that simple architectures like Basic CNN 1 can outperform more complex models under certain conditions. The importance of preprocessing techniques, like histogram equalization, was also highlighted for improving model performance.
This project was completed as part of the BIM453 course. Special thanks to Data Source for providing the dataset.
This project is licensed under the MIT License - see the LICENSE file for details.
Note: Replace path/to/... with actual paths to the images or outputs generated by your code.