SSDiff: A Contrast-Free Virtual LGE Generator for Acute Myocardial Infarction with Joint Segmentation via Diffusion...

Area of Science:

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

Background:

• Myocardial infarction (MI) is a leading cause of mortality, with late gadolinium enhancement (LGE) cardiac MRI being the gold standard for assessing myocardial viability.
• Current LGE techniques require contrast agents, complex protocols, and incur additional costs, limiting their widespread clinical use.
• Existing virtual LGE methods often overlook T2-weighted short-tau inversion recovery (T2-STIR) sequences, crucial for detecting acute MI edema, and typically need manual infarct delineation.

Purpose of the Study:

• To develop and validate SSDiff, a multitask conditional diffusion framework for synthesizing contrast-free virtual LGE images from routine cine and T2-STIR cardiac MRI.
• To enable simultaneous segmentation of myocardium, ventricular blood pool, and infarct regions using the proposed framework.
• To provide a clinically feasible and cost-effective alternative for acute myocardial infarction assessment.

Main Methods:

• SSDiff employs a multitask conditional diffusion model integrating cine and T2-STIR sequences for virtual LGE synthesis.
• A feature-disentangled attention module isolates sequence-specific information to guide the diffusion process.
• A cross-fusion module aligns synthesis and segmentation decoders, enabling mutual optimization for improved accuracy.

Main Results:

• SSDiff demonstrated significant improvements in synthetic virtual LGE image quality compared to existing methods.
• The framework achieved superior accuracy in segmenting myocardium, blood pool, and infarct regions.
• Evaluation on a multi-center, multi-vendor dataset of 409 subjects confirmed the robustness and generalizability of SSDiff.

Conclusions:

• SSDiff offers a promising, contrast-free approach for assessing myocardial viability and infarct extent in acute MI.
• The generated paired image-mask samples can augment training datasets for AI models, addressing scarcity of LGE-scarred training data.
• SSDiff holds significant practical utility and translational potential as a clinically feasible alternative to conventional LGE-MRI.