GI Tract Image Segmentation

Overview

This project implements a U-Net deep learning model for medical image segmentation, specifically targeting segmentation of the gastrointestinal (GI) tract from medical scans. The pipeline includes custom data preprocessing, a modular training process, and exportable predictions in Run-Length Encoding (RLE) format.

Key Features

• Data Pipeline:
• Custom data generator to handle large datasets with RLE-encoded masks.
• Dynamic resizing of images and masks to a configurable target size.

• Flexible test mode for visualizing individual predictions and ground truths.

• Model Architecture:

• Based on TransUNet, combining convolutional layers with transformer-based features for superior performance.

• Option to load pre-trained weights for transfer learning or train from scratch.

• Evaluation:

• Metrics include Dice coefficient, accuracy, and visual analysis of predictions.

• Visualization overlays for ground truths and predictions on original images.

• Export:

• Saves predictions in RLE format compatible with Kaggle competitions or downstream pipelines.

Requirements

To set up and run the project, ensure the following dependencies are installed:
- TensorFlow 2.8+
- Keras
- NumPy
- Pandas
- OpenCV
- Matplotlib
- Scikit-learn

Usage

1. Running the Pipeline

Run the main script to train, evaluate, or export predictions:
python GI-Tract-Image-Segmentation.py

2. Training the Model

If training from scratch:
- Automatically splits the data into training, validation, and test sets.
- Implements early stopping, learning rate scheduling, and model checkpointing.
- Saves the best model weights to the output/ directory.

3. Evaluating the Model

During evaluation, the script:
- Computes Dice coefficient and loss for each test sample.
- Visualizes predictions and overlays with ground truths.

Data Pipeline

The project includes a highly modular pipeline:
- Custom Generator:
- Decodes RLE masks into binary masks dynamically.
- Handles resizing, augmentation, and batch generation.
- Training Pipeline:
- Modularized for scalability and customization.
- Includes checkpoints and CSV logging of training metrics.

Model Architecture

• Base Model: TransUNet
• Combines transformer layers for long-range dependency capture with CNNs for spatial feature extraction.
• Custom Modifications:
• Configurable input size.
• Optional pre-trained weights for transfer learning.