CT-Heart-Segmentation

Overview

This project focuses on automated heart segmentation from CT scans using deep learning. We applied data augmentation techniques using the Albumentations library to enhance model generalization and trained a U-Net model to segment heart structures accurately.

Medical image segmentation is crucial in cardiovascular disease diagnosis, treatment planning, and surgical interventions. This project aims to provide a robust and efficient deep learning pipeline for heart segmentation from CT images.

Dataset

The dataset consists of CT scan images with corresponding segmentation masks representing heart structures. The images undergo preprocessing and augmentation before being fed into the segmentation model.

Kaggle link for Dataset: https://www.kaggle.com/datasets/nikhilroxtomar/ct-heart-segmentation

Preprocessing Steps:

✅ Resizing images to a fixed resolution
✅ Normalizing pixel values for stable training

Data Augmentation using Albumentations

To improve the model's generalization and performance, we applied data augmentation using the Albumentations library. This helps to create diverse training samples, reducing overfitting and improving robustness.

Augmentations Applied:

✅ RandomBrightnessContrast → Adjusts brightness and contrast randomly to simulate variations in scanning conditions
✅ HueSaturationValue → Modifies the hue, saturation, and value of the image to introduce color variations
✅ RGBShift → Shifts the red, green, and blue channels to create different color variations
✅ RandomGamma → Randomly adjusts gamma levels to enhance or darken image regions
✅ CLAHE (Contrast Limited Adaptive Histogram Equalization) → Enhances local contrast to improve visibility of structures
✅ ChannelShuffle → Randomly shuffles color channels to increase diversity in training samples

Model Architecture: U-Net for Segmentation

We implemented a U-Net model, a well-known CNN-based architecture for biomedical image segmentation. U-Net is effective in capturing both local and global spatial information using its encoder-decoder structure with skip connections.

U-Net Architecture Highlights:

🔹 Encoder (Contracting Path) → Captures spatial features using convolutional layers
🔹 Bottleneck → Connects the encoder and decoder with high-level feature maps
🔹 Decoder (Expanding Path) → Recovers spatial resolution for precise segmentation
🔹 Skip Connections → Retains fine-grained details by merging encoder features

**Training Process **

The model was trained using TensorFlow/Keras with the following setup:
✅ Loss Function → Dice Loss
✅ Optimizer → Adam optimizer for efficient learning
✅ Learning Rate Scheduling → Reduces learning rate on plateaus
✅ Evaluation Metrics → Dice Coefficient, IoU, Accuracy