SSL-DA: Semi-and Self-Supervised Learning with Dual Attention for Echocardiogram Segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence in cardiology
• Image segmentation

Background:

• Echocardiogram analysis is vital for diagnosing heart disease.
• Accurate left ventricle (LV) segmentation in echocardiograms is challenging and time-consuming.
• Existing automated methods lack accuracy and reproducibility; manual methods are inefficient.

Purpose of the Study:

• To introduce a novel semi- and self-supervised learning with dual attention (SSL-DA) framework for echocardiogram segmentation.
• To improve the accuracy and efficiency of left ventricle (LV) segmentation.
• To provide a robust tool for clinical application in cardiac diagnostics.

Main Methods:

• Utilized a temporal masking network for pre-training to capture echocardiogram periodicity and optimize initialization.
• Employed a semi-supervised network with channel and spatial attention mechanisms for LV segmentation.
• Evaluated the SSL-DA framework on the EchoNet-Dynamic and CAMUS datasets.

Main Results:

• Achieved a Dice similarity coefficient of 93.34% on the EchoNet-Dynamic dataset, outperforming prior CNN-based models.
• Demonstrated strong generalization ability through ablation experiments on the CAMUS dataset.
• Confirmed rapid and accurate LV segmentation capabilities of the SSL-DA framework.

Conclusions:

• The SSL-DA framework offers a significant advancement in automated echocardiogram segmentation.
• SSL-DA shows potential for robust and efficient clinical application in diagnosing heart conditions.
• This method addresses limitations of current segmentation techniques, improving cardiac assessment.