CMR-BENet: A confidence map refinement boundary enhancement network for left ventricular myocardium segmentation

Area of Science:

• Medical imaging analysis
• Cardiovascular image processing
• Artificial intelligence in medicine

Background:

• Left ventricular myocardium segmentation is crucial for clinical diagnosis and prognosis.
• Image quality issues like noise and motion hinder accurate segmentation.
• Existing methods struggle with boundary accuracy and reliability assessment.

Purpose of the Study:

• To develop an end-to-end network for improved left ventricular myocardium segmentation.
• To enhance segmentation by integrating contextual information with edge structure and confidence maps.
• To address limitations in current methods regarding boundary detection and prediction reliability.

Main Methods:

• Proposed the confidence map refinement boundary enhancement network (CMR-BENet).
• Incorporated a layer semantic-aware module (LSA) for adaptive feature fusion.
• Integrated an edge information enhancement module (EIE) and a confidence map-based refinement module (CMR) for boundary and structure awareness.

Main Results:

• CMR-BENet achieved superior performance in left ventricular myocardium segmentation across multiple datasets.
• Demonstrated high Dice scores (e.g., 87.71%, 79.33%, 89.11%) on echocardiography and cardiac MRI data.
• Outperformed existing segmentation methods in challenging, noisy imaging conditions.

Conclusions:

• Edge information and learnable confidence maps effectively characterize myocardial shape and structure.
• The proposed method refines segmentation results and assesses prediction reliability.
• Findings offer valuable support for physicians in cardiac diagnosis and treatment planning.