Convolutional Neural Network With Shape Prior Applied to Cardiac MRI Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Cardiovascular Imaging

Background:

• Accurate segmentation of cardiac magnetic resonance images (CMRI) is crucial for diagnosing cardiovascular diseases.
• Existing methods often require manual preprocessing or are computationally intensive.

Purpose of the Study:

• To develop a novel convolutional neural network (CNN) architecture for automated segmentation of cardiac structures in CMRI.
• To improve the accuracy and efficiency of segmenting left and right ventricle contours.

Main Methods:

• A U-net based CNN architecture was extended with a cardiac shape prior and a tailored loss function.
• The model processes raw CMRI slices, requiring no manual preprocessing or cropping.
• A multiresolution grid architecture enables learning of both high and low-level features for accurate localization.

Main Results:

• The model achieved an average Dice coefficient of [Formula: see text] and an average 3-D Hausdorff distance of [Formula: see text] mm on the ACDC-MICCAI 2017 dataset.
• Segmentation of multi-slice CMRI (ventricle contours) was performed rapidly, in 0.18 seconds.
• The system accurately segmented the endocardium and epicardium of the left ventricle, and the endocardium of the right ventricle.

Conclusions:

• The proposed CNN architecture offers an efficient and accurate solution for cardiac image segmentation.
• The integration of a cardiac shape prior and specialized loss function enhances segmentation performance.
• This automated approach has the potential to streamline cardiovascular image analysis workflows.