Brain Tumor Prediction using ResNet Architecture

This repository contains a deep learning project for predicting brain tumors using CT and MRI images. The model leverages a custom-designed ResNet architecture implemented using TensorFlow Core API. The training pipeline includes a custom training loop and distributed GPU training for efficiency and scalability.

Key Features

1. ResNet Architecture: The model architecture is based on the ResNet (Residual Network) design, which is well-suited for extracting complex patterns from medical imaging data.

2. TensorFlow Core API: The architecture is built from scratch using TensorFlow Core API for maximum control and flexibility.

3. Custom Training Loop: A custom training loop was implemented to provide fine-grained control over the training process, loss calculation, and metrics tracking.

4. Distributed GPU Training: To handle large datasets and expedite training, distributed training across multiple GPUs was employed using TensorFlow's distribution strategies.

Dataset

The dataset consists of labeled CT and MRI brain images categorized into two classes:

1. Tumor

2. Non-Tumor

Preprocessing Steps:

1. Resized all images to a fixed size of 224x224 pixels.

2. Normalized pixel values to the range [0, 1].

Model Architecture

The ResNet architecture used in this project comprises the following components:

1. Input Layer: Accepts 224x224x3 image tensors.

2. Residual Blocks: Stacked blocks of convolutional layers with skip connections to alleviate the vanishing gradient problem.

3. Global Average Pooling: Reduces the spatial dimensions while retaining key features.

4. Dense Layers: Fully connected layers with ReLU activation.

5. Output Layer: Single neuron with sigmoid activation for binary classification.

Implementation Details

Dependencies

• TensorFlow

• NumPy

• Matplotlib

• scikit-learn

Training Pipeline

• Model Definition: The ResNet architecture was defined using TensorFlow Core API.

• Loss Function: Binary cross-entropy loss was used to quantify the difference between predicted and actual labels.

• Optimizer: Adam optimizer with a learning rate scheduler for dynamic learning rate adjustments.

• Metrics: Accuracy were tracked during training.

• Custom Training Loop: Included forward propagation, backpropagation, and metrics logging for each batch.

• Distributed Training: TensorFlow's tf.distribute.MirroredStrategy was used to enable multi-GPU training.

Training Results

• Training Accuracy: Achieved >90% accuracy after 80 epochs.

• Validation Accuracy: Achieved > 90%.

Acknowledgments

1. Open-source frameworks like TensorFlow and NumPy.

2. Publicly available datasets for medical imaging research. Dataset Link:- DataSet

3. Community support and resources that facilitated the development of this project.

License

This project is licensed under the MIT License. See the LICENSE file for details.