Clinically Relevant Myocardium Segmentation in Cardiac Magnetic Resonance Images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Background:

• Deep learning excels at cardiac MR (CMR) image segmentation but struggles with irregularities like contour breaks.
• Manual correction of segmentation by clinicians is time-consuming and impacts myocardium condition evaluation.
• Existing methods lack robustness for clinical decision support systems due to segmentation defects.

Purpose of the Study:

• To develop a deep learning system that handles irregularities in myocardium segmentation.
• To impose structural constraints on existing segmentation outputs for clinical validity.
• To create a more automated and reliable system for cardiac image analysis.

Main Methods:

• A two-stage deep neural network pipeline was proposed: initial segmentation followed by a refinement network.
• The refinement network imposes structural constraints to correct segmentation defects.
• The system was tested on datasets from four diverse sources.

Main Results:

• The refinement model consistently improved segmentation outputs across different initial networks.
• Significant improvements observed: up to 8% in Dice Coefficient and 18 pixels in Hausdorff Distance.
• Qualitative and quantitative performance enhancements were noted for all tested segmentation networks.

Conclusions:

• The proposed refinement strategy effectively addresses segmentation irregularities in CMR images.
• This work is a significant step towards fully automatic myocardium segmentation systems.
• The methodology is potentially generalizable to other segmentation tasks with regular structures.