Cross-Modality LGE-CMR Segmentation Using Image-to-Image Translation Based Data Augmentation.

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Accurate segmentation of ventricle and myocardium in late gadolinium enhancement cardiac magnetic resonance (LGE-CMR) is vital for myocardial infarction (MI) analysis.
• The complex patterns in LGE-CMR and scarcity of labeled data hinder automatic segmentation development.
• Existing methods often require extensive manually annotated LGE-CMR datasets.

Purpose of the Study:

• To develop an unsupervised LGE-CMR segmentation algorithm that overcomes the limitations of labeled data scarcity.
• To enhance the accuracy and efficiency of automatic segmentation for LGE-CMR analysis.
• To leverage data augmentation techniques for improved performance in medical image segmentation.

Main Methods:

• Proposed an unsupervised segmentation algorithm utilizing multiple style transfer networks for data augmentation.
• Employed two distinct style transfer networks to adapt readily available annotated balanced-Steady State Free Precession (bSSFP)-CMR images.
• Generated synthetic LGE-CMR images via style transfer for training an improved U-Net model.

Main Results:

• The developed algorithm successfully segments LGE-CMR images without requiring labeled LGE-CMR data.
• Validation experiments confirmed the effectiveness and advantages of the proposed unsupervised approach.
• The style transfer augmentation strategy proved beneficial for training the segmentation model.

Conclusions:

• The proposed unsupervised LGE-CMR segmentation method effectively addresses the challenge of limited labeled data.
• Style transfer-based data augmentation is a viable strategy for improving medical image segmentation tasks.
• This algorithm offers a promising solution for more accessible and accurate MI analysis using LGE-CMR.