ScarNet: A Novel Foundation Model for Automated Myocardial Scar Quantification from LGE in Cardiac MRI

Overview

ScarNet is a deep learning model specifically designed for automated myocardial scar segmentation and quantification in Late Gadolinium Enhancement (LGE) Cardiac MRI. By leveraging a hybrid architecture that integrates MedSAM's Vision Transformer (ViT)-based encoder with a U-Net decoder, ScarNet achieves state-of-the-art performance in scar segmentation and quantification.

Features

1. Hybrid Architecture:

   • Combines the global context awareness of transformers with the precise localization capabilities of CNNs.
   • MedSAM's ViT-based encoder for global feature extraction
   • U-Net decoder for precise localization
   • Incorporates specialized attention mechanisms, including SE layers and ScarAttention blocks.

2. Innovative Loss Function:

   • Combines Dice, Focal, and Cross-Entropy losses to address class imbalance and improve segmentation accuracy.

3. Robust Performance:

   • Demonstrates high Dice scores (median = 0.912) for scar segmentation, outperforming existing models.
   • Robust against noise perturbations, ensuring reliability in varying imaging conditions.

4. Efficient Training Pipeline:

5. Image preprocessing and augmentation

6. Parallel processing:

   • MedSAM pathway
   • U-Net pathway

7. Adaptive feature fusion

8. Output generation

9. Efficient Inference Pipeline:

   • Designed for seamless integration into clinical workflows.

Key Results

• ScarNet achieves a median scar Dice score of 0.912, compared to 0.046 for MedSAM and 0.638 for nnU-Net.
• Demonstrates exceptional robustness across different image qualities and noise conditions.

Figures

Model Architecture

Figure 1: Hybrid architecture combining MedSAM's ViT-based with U-Net's multi-scale decoder. This figure highlights ScarNet's integration of attention mechanisms for precise scar segmentation.

Figure 2: Detailed encoder-decoder overview, showcasing attention blocks and SE layers that enhance feature extraction and refine scar segmentation.

Comparative Analysis

Figure 3: Comparative visualization of ScarNet, MedSAM, and nnU-Net segmentations. ScarNet demonstrates superior segmentation accuracy across different cardiac regions.

Figure 4: Performance analysis across varying training data sizes. ScarNet consistently outperforms MedSAM and nnU-Net in both myocardium and scar segmentation.

Figure 5: Representative test cases demonstrating ScarNet's segmentation performance compared to MedSAM and nnU-Net.

Figure 6: Performance metrics for scar volume quantification, including Bland-Altman plots and correlation analyses. ScarNet achieves near-perfect correlation with manual segmentation.

Methodology

Dataset

• Retrospective analysis using de-identified LGE images from 736 patients with ischemic cardiomyopathy.
• Training set: 552 patients; Testing set: 184 patients.

Training Pipeline

1. Image preprocessing and augmentation.
2. Parallel processing through MedSAM and U-Net pathways.
3. Adaptive feature fusion and output generation.

Loss Function

The multi-component loss function is defined as: [ \mathcal{L}_{ScarNet} = \lambda_1 \text{FTL} + \lambda_2 \text{DL} + \lambda_3 \text{CE} ]

• FTL: Focal Tversky Loss
• DL: Dice Loss
• CE: Cross-Entropy Loss

Installation and Usage

Prerequisites

• Python 3.8+
• PyTorch 1.9+
• CUDA-capable GPU

Setup

1. Clone the repository:

git clone https://github.com/NedaTavakoli/ScarNet.git
cd ScarNet

2. Install dependencies:

pip install -r requirements.txt

3. Run training:

python train.py

# With custom config
python train.py --config path/to/config.yaml

Data Preparation

Organize your data in the following structure:

data/
├── training/
│   ├── images/
│   └── masks/
└── testing/
    ├── images/
    └── masks/

Training Configuration

Modify config.py to adjust:

• Learning rate
• Batch size
• Training epochs
• Model architecture parameters
• Augmentation settings

License

MIT License - See LICENSE file for details

Contact

Neda Tavakoli
Email: neda.tavakoli@northwestern.edu

Citation

@article{ScarNet2024,
    title={ScarNet: A Novel Foundation Model for Automated Myocardial Scar Quantification from LGE},
    author={Tavakoli, Neda and others},
    year={2024}
}