# DeepLearning-HeartSegmentation

# DeepLearning-HeartSegmentation

Heart segmentation in chest CT scans 

Heart segmentation in chest CT scans 

## Introduction

## Introduction

This project aims to develop and evaluate deep learning models for the accurate segmentation of the heart in medical images, particularly in non-contrast and non-gated computed tomography (CT) scans. The precise delineation of the heart region is crucial for various medical applications.

This project aims to develop and evaluate deep learning models for the accurate segmentation of the heart in medical images, particularly in non-contrast and non-gated computed tomography (CT) scans. The precise delineation of the heart region is crucial for various medical applications.

## Models and Architectures

## Models and Architectures

### U-Net++

### U-Net++

We employed the U-Net++ architecture as a base model for heart segmentation. The model was trained using a Binary Cross-Entropy loss function. After 100 epochs of training, the model achieved an average Dice similarity coefficient of 0.72 and a Jaccard index of 0.67 on the validation set.

We employed the U-Net++ architecture as a base model for heart segmentation. The model was trained using a Binary Cross-Entropy loss function. After 100 epochs of training, the model achieved an average Dice similarity coefficient of 0.72 and a Jaccard index of 0.67 on the validation set.

### MA-Net

### MA-Net

The Multi-scale Attention Net (MA-Net) architecture was implemented to improve heart segmentation. After 100 epochs, this model outperformed U-Net++ with an average Dice coefficient of 0.85 and a Jaccard index of 0.82 on the validation set. Further training for 200 epochs reduced oscillations in the training and validation metrics.

The Multi-scale Attention Net (MA-Net) architecture was implemented to improve heart segmentation. After 100 epochs, this model outperformed U-Net++ with an average Dice coefficient of 0.85 and a Jaccard index of 0.82 on the validation set. Further training for 200 epochs reduced oscillations in the training and validation metrics.

<p align='center'>

<p align='center'>

<img src='https://github.com/dcrovo/DeepLearning-HeartSegmentation/blob/main/imgs/MANET_TRAIN.png' />  

<img src='https://github.com/dcrovo/DeepLearning-HeartSegmentation/blob/main/imgs/MANET_TRAIN.png?raw=true' />  

<img src='https://github.com/dcrovo/DeepLearning-HeartSegmentation/blob/main/imgs/MANET_DICE_VAL.png' />  

<img src='https://github.com/dcrovo/DeepLearning-HeartSegmentation/blob/main/imgs/MANET_DICE_VAL.png?raw=true' />  

</p>

</p>

## Results

## Results

This is a sample CT scan segmented by the MANET model in red and by a certified board radiologist in green.

This is a sample CT scan segmented by the MANET model in red and by a certified board radiologist in green.

<p align='center'>

<p align='center'>

<img src='https://github.com/dcrovo/DeepLearning-HeartSegmentation/blob/main/imgs/final.png' width=496 height=515/>  

<img src='https://github.com/dcrovo/DeepLearning-HeartSegmentation/blob/main/imgs/final.png?raw=true' width=496 height=515/>  

</p>

</p>

In summary, the MA-Net model demonstrated superior performance in heart segmentation compared to U-Net++. It achieved a Dice coefficient of 0.883, which compares favourably with state-of-the-art models. However, both models exhibited some overfitting, likely due to the limited size of the training dataset.

In summary, the MA-Net model demonstrated superior performance in heart segmentation compared to U-Net++. It achieved a Dice coefficient of 0.883, which compares favourably with state-of-the-art models. However, both models exhibited some overfitting, likely due to the limited size of the training dataset.

## Future Work

## Future Work

To enhance the models further, future work will focus on hyperparameter optimization and increased data augmentation. The well-segmented heart regions will be used as a crucial step for addressing broader medical image analysis challenges.

To enhance the models further, future work will focus on hyperparameter optimization and increased data augmentation. The well-segmented heart regions will be used as a crucial step for addressing broader medical image analysis challenges.