Semi-Automatic Refinement of Myocardial Segmentations for Better LVNC Detection

Area of Science:

• Medical Imaging
• Cardiology
• Artificial Intelligence

Background:

• Accurate left ventricular myocardium segmentation in cardiac MRI is crucial for diagnosing left ventricular non-compaction cardiomyopathy (LVNC).
• Existing segmentation databases require enhancement for reliable deep learning model training.
• This study addresses the need for improved myocardial segmentation quality in cardiac MRI for LVNC diagnosis.

Purpose of the Study:

• To develop and validate a semi-automatic framework for refining cardiac MRI segmentation.
• To improve the quality of myocardial segmentation datasets for training deep learning models.
• To enhance the accuracy of diagnosing left ventricular non-compaction cardiomyopathy (LVNC).

Main Methods:

• A semi-automatic framework combining neural network outputs with expert corrections.
• Implementation of a blob-selection method to rectify segmentation errors and neural network hallucinations.
• Utilization of a cross-validation process with a baseline U-Net model for segmentation refinement.

Main Results:

• The proposed methods demonstrated improved segmentation accuracy across multi-hospital datasets.
• The blob-selection technique significantly boosted the Dice coefficient for the Trabecular Zone by up to 0.06 in specific populations.
• Enhanced segmentation quality was achieved, particularly in challenging regions of the myocardium.

Conclusions:

• The developed semi-automatic framework effectively enhances the quality of cardiac MRI segmentation datasets.
• This improved dataset provides a more robust foundation for developing accurate deep learning models for LVNC diagnosis.
• The study highlights the potential of refined segmentation techniques in advancing cardiovascular disease diagnostics.