Feature Tensor Low-Rank Representation Network for Semi-Supervised Echocardiography Video Left Ventricle Segmentation

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Artificial Intelligence in Medicine

Background:

• Echocardiography is vital for non-invasive cardiac assessment, but accurate segmentation of the left ventricle is challenging.
• Technical limitations include speckle noise, low contrast, and incomplete annotations, hindering precise analysis of cardiac structure and function.
• Advanced image processing techniques are needed to overcome these hurdles in echocardiography video analysis.

Purpose of the Study:

• To develop a novel semi-supervised deep learning method for accurate left ventricular segmentation in echocardiography videos.
• To address challenges like noise and low contrast by leveraging low-rank reconstruction of video feature tensors.
• To improve the understanding of cardiac structure and function through enhanced segmentation.

Main Methods:

• A shared 2D convolutional backbone extracts deep semantic features from echocardiography frames.
• Features are temporally stacked into high-order feature tensors, exploiting spatiotemporal correlations.
• Tensor singular value thresholding in the transform domain is used for compact feature representation and noise reduction.

Main Results:

• The proposed semi-supervised method achieved superior left ventricle segmentation accuracy on the CAMUS dataset.
• Experimental results demonstrated that the technique outperforms existing state-of-the-art methods.
• The method effectively reduces feature redundancy and enhances segmentation precision.

Conclusions:

• The developed deep learning methodology offers a promising approach for left ventricle segmentation in echocardiography.
• The superior performance indicates significant potential for clinical applications in cardiac diagnostics.
• This technique can advance the accuracy and reliability of echocardiography-based heart health assessments.