A Novel Framework With Weighted Decision Map Based on Convolutional Neural Network for Cardiac MR Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Accurate segmentation of cardiac structures in magnetic resonance imaging (MRI) is crucial for diagnosing cardiovascular diseases.
• Existing deep learning methods face challenges with over- and under-segmentation in short-axis view cardiac MRI.
• Addressing these segmentation inaccuracies is vital for reliable cardiovascular assessments.

Purpose of the Study:

• To develop a novel two-stage deep learning framework for simultaneous segmentation of myocardium, left ventricle, and right ventricle from cardiac MRI.
• To overcome common segmentation errors like over- and under-segmentation in short-axis views.
• To improve the accuracy of cardiac structure segmentation across different MRI sequences.

Main Methods:

• A two-stage framework incorporating a decision map extractor (cascaded U-Net++) and a cardiac segmenter (MDFA-Net).
• The decision map extractor generates pixel-wise category predictions to guide segmentation.
• The cardiac segmenter utilizes multiscale dual-path feature aggregation for enhanced segmentation accuracy.

Main Results:

• The proposed method achieved average Dice coefficients of 84.70% for myocardium, 86.00% for left ventricle, and 86.31% for right ventricle on the MyoPS 2020 dataset.
• Demonstrated effectiveness in segmenting key cardiac structures simultaneously.
• Successfully addressed over- and under-segmentation issues in cardiac MRI.

Conclusions:

• The novel two-stage framework provides accurate and robust segmentation of cardiac structures from MRI.
• This approach offers a significant advancement for cardiovascular disease diagnosis and analysis.
• The method shows promise for clinical application in cardiac MRI segmentation.