An Efficient Capsule-based Network for 2D Left Ventricle Segmentation in Echocardiography Images

Area of Science:

• Cardiovascular Imaging Analysis
• Deep Learning in Medical Diagnostics
• Biomedical Image Segmentation

Background:

• Accurate segmentation of cardiac chambers, particularly the left ventricle (LV), is crucial for diagnosing and managing cardiovascular diseases.
• Manual segmentation of LV in echocardiograms is time-consuming and prone to inaccuracies due to image quality limitations.
• While Convolutional Neural Networks (CNNs) are popular for medical image segmentation, they have limitations in preserving spatial information and require extensive datasets and parameters.

Purpose of the Study:

• To propose an optimized capsule-based network, SegCaps, for accurate left ventricle (LV) segmentation in echocardiographic images.
• To address the limitations of CNNs, such as ignoring spatial information and requiring large datasets and parameters.
• To evaluate the performance of SegCaps against the standard 2D-UNet on the CAMUS dataset for LV segmentation.

Main Methods:

• Development and application of an optimized capsule network (SegCaps) for object segmentation.
• Utilized the CAMUS dataset for training and evaluating the SegCaps model on echocardiographic images.
• Comparative analysis of SegCaps against the 2D-UNet model, focusing on segmentation accuracy and parameter efficiency.

Main Results:

• The SegCaps network achieved an average Dice Similarity Coefficient (DSC) of 84.48% for LV segmentation, outperforming the 2D-UNet's 83.28%.
• SegCaps demonstrated a 1.44% improvement in DSC compared to 2D-UNet.
• The proposed SegCaps method utilized 92.77% fewer parameters than 2D-UNet, indicating significant efficiency gains.

Conclusions:

• The optimized SegCaps network provides accurate and efficient left ventricle (LV) segmentation in echocardiographic images.
• Capsule networks offer an advantage over traditional CNNs by preserving spatial information and requiring fewer parameters.
• This method enables more precise clinical evaluations of cardiac parameters like ejection fraction and ventricular volumes.