Automatic left ventricle segmentation from cardiac magnetic resonance images using a capsule network

Area of Science:

• Cardiovascular Imaging and Analysis
• Medical Image Segmentation
• Deep Learning in Healthcare

Background:

• Accurate quantification of cardiac function relies on precise segmentation of the left ventricle (LV) from cardiac magnetic resonance imaging (MRI).
• Traditional manual LV segmentation is labor-intensive and susceptible to inter-observer variability.
• Automated methods are crucial for efficient and reproducible cardiac function assessment.

Purpose of the Study:

• To develop and evaluate a novel capsule network-based automated method for segmenting the left ventricle (LV) in cardiac MRI.
• To improve the accuracy and efficiency of LV segmentation compared to traditional manual approaches.
• To enable more reliable quantification of cardiac volumetric functions.

Main Methods:

• A capsule network architecture was employed for precise LV segmentation, utilizing Fourier analysis and circular Hough transform for initial LV localization.
• The capsule network replaced traditional pooling layers with convolutional strides and dynamic routing to preserve input information.
• Segmentation results were refined using postprocessing techniques, including threshold segmentation and morphological operations.

Main Results:

• The proposed capsule network method achieved high segmentation accuracy, with Dice scores of 0.922±0.05 for end-diastolic and 0.898±0.11 for end-systolic phases on the ACDC 2017 dataset.
• Quantitative and visual comparisons demonstrated superior segmentation performance compared to traditional methods.
• The algorithm effectively segmented the LV from cardiac MRI, confirming its practical utility.

Conclusions:

• The developed capsule network-based method, combined with postprocessing, significantly enhances segmentation accuracy in cardiac MRI.
• This study represents the first systematic evaluation of a deep learning capsule network for end-to-end LV segmentation.
• The findings suggest a promising automated approach for accurate cardiac function assessment using MRI.