JOINT DEEP LEARNING FOR IMPROVED MYOCARDIAL SCAR DETECTION FROM CARDIAC MRI

Area of Science:

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

Background:

• Automated myocardial scar identification from late gadolinium enhancement cardiac magnetic resonance images (LGE-CMR) is challenged by image noise and artifacts.
• Existing methods often treat scar detection and myocardium segmentation as separate tasks, limiting performance.

Purpose of the Study:

• To develop a novel joint deep learning (JDL) framework for improved automated myocardial scar detection in LGE-CMR.
• To leverage simultaneously learned myocardium segmentations to mitigate the impact of noise and artifacts.

Main Methods:

• A novel JDL framework incorporating a message passing module was designed.
• Myocardium segmentation information is directly passed to guide scar detectors within the network.
• The framework jointly learns segmentation and scar detection tasks for enhanced information exploitation.

Main Results:

• The JDL framework demonstrated effectiveness in automated left ventricular (LV) scar detection on LGE-CMR images.
• Experimental results showed superior performance compared to state-of-the-art methods, including two-step and multitask learning approaches.
• The JDL approach outperformed methods with indirect task interaction.

Conclusions:

• The proposed JDL framework offers a significant advancement in automated myocardial scar identification from LGE-CMR.
• This method holds potential for improved risk prediction in heart disease patients and enhanced response prediction for cardiac resynchronization therapy (CRT).
• The joint learning approach effectively utilizes complementary information from segmentation to benefit scar detection.