/BrainTumorDetectionResearch

Primary LanguageJupyter Notebook

Brain Tumor Detection Using MRI Scans with CAM-Guided Attention

Overview

This repository contains the implementation of a deep learning model designed for brain tumor detection using Magnetic Resonance Imaging (MRI) scans. The model utilizes Class Activation Maps (CAM) to provide interpretable insights and guide channel attention, enabling accurate and explainable predictions.

Features

  • Binary Classification: Tumor vs. No Tumor.
  • Interpretable Predictions: CAM visualizations highlight critical regions in MRI scans for model decisions.
  • Dataset Preprocessing: Cropping, resizing, and normalization techniques are used to enhance model performance.
  • Attention Mechanism: CAM-guided attention improves feature extraction and model generalization.

Table of Contents

Dataset

The model was trained and evaluated using a dataset of 2,065 MRI images (after augmentation) sourced from:

Data Preprocessing

  1. Cropping: Removes irrelevant background information.
  2. Resizing: Standardizes all images to a fixed size of 240x240.
  3. Normalization: Scales pixel values to [0, 1] for numerical stability.
  4. Augmentation: Increases dataset size to improve model generalization.

Model Architecture

Layer Details Output Shape
Input MRI image (3 channels, 240x240) (3, 240, 240)
Conv2D 32 filters, kernel size=5, stride=1, padding=2 (32, 240, 240)
BatchNorm2D Batch normalization on 32 channels (32, 240, 240)
ReLU Activation Introduces non-linearity (32, 240, 240)
MaxPool2D Kernel size=4, stride=4 (32, 60, 60)
CAM-Guided Attention Refines feature maps (32, 60, 60)
MaxPool2D Kernel size=4, stride=4 (32, 15, 15)
Flatten Converts features into 1D vector (32 * 15 * 15)
Fully Connected Sigmoid activation for binary classification (1)

Training Pipeline

  • Optimizer: Adam with a learning rate of 0.001.
  • Loss Function: Binary Cross Entropy Loss (BCELoss) for probability-based classification.
  • Progressive Attention: CAM-guided attention is introduced after two epochs to stabilize training.
  • Metrics Tracked:
    • Training Loss
    • Validation Loss
    • Training Accuracy
    • Validation Accuracy

Results and Visualizations

Performance Metrics

  • Validation Accuracy: ~98.12%
  • Test Accuracy: ~94.87%

How to Use

Clone Repository

git clone https://github.com/your-username/brain-tumor-detection.git
cd brain-tumor-detection

Run the python notebook on google colab or jupyter

Future Work

  1. Extend to multiclass classification (e.g., distinguishing between benign and malignant tumors).
  2. Integrate additional interpretability techniques like Grad-CAM++.
  3. Optimize the model for real-time inference on lightweight devices.

Acknowledgments

  • CAM Research: Inspired by Woo et al., 2018.
  • Dataset Contributors: Kaggle community for publicly available MRI datasets.