CMMCSegNet: Cross-Modality Multicascade Indirect LGE Segmentation on Multimodal Cardiac MR.

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Artificial Intelligence in Medicine

Background:

• Late-Gadolinium Enhancement (LGE) in cardiac magnetic resonance (CMR) is crucial for visualizing myocardial infarction and assessing cardiac viability.
• Accurate segmentation of LGE is challenging due to limited labeled data and low contrast in CMR images.
• Multimodal CMR segmentation is vital for clinical diagnosis but automated methods face significant hurdles.

Purpose of the Study:

• To develop an automated and accurate method for cardiac LGE segmentation in CMR images.
• To address the challenge of limited labeled LGE data by utilizing cross-modality learning.
• To enhance the assessment of myocardial viability and clinical diagnosis through improved LGE segmentation.

Main Methods:

• A cross-modality multicascade framework combining a translation network and a segmentation network was proposed.
• A novel multicascade pix2pix network was designed to segment a synthetically generated (fake) bSSFP sequence.
• Perceptual loss, using features from a pre-trained VGG network, was incorporated for improved segmentation accuracy.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art approaches on a multimodal CMR dataset.
• Evaluation across different network structures and adversarial losses confirmed the method's effectiveness.
• The approach achieved high dice accuracy in indirect cardiac LGE segmentation.

Conclusions:

• The developed cross-modality framework offers a promising solution for indirect cardiac LGE segmentation.
• This technique can overcome limitations posed by scarce labeled LGE data in CMR.
• The method holds potential for significant clinical applications in diagnosing and managing cardiac conditions.