An improved contrastive learning network for semi-supervised multi-structure segmentation in echocardiography

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiac diseases pose a significant mortality risk.
• Echocardiography is a vital, non-invasive diagnostic tool.
• Accurate segmentation of cardiac structures is essential but challenging due to image quality and anatomical variations.

Purpose of the Study:

• To develop a semi-supervised method for precise cardiac structure segmentation in echocardiographic images.
• To address challenges like low contrast, incomplete structures, and unclear borders.
• To leverage unlabeled data for improved segmentation performance.

Main Methods:

• Applied a contrastive learning strategy.
• Developed a semi-supervised learning approach for echocardiographic image segmentation.
• Evaluated the method on the public CAMUS dataset.

Main Results:

• Achieved high Dice Similarity Coefficients (DSC) on two-chamber (2CH) and four-chamber (4CH) images, even with limited labeled data (e.g., 0.916 and 0.928 with full labels).
• Demonstrated superior performance and fewer parameters compared to existing methods.
• Effectively improved segmentation accuracy despite image quality challenges.

Conclusions:

• The proposed semi-supervised method enhances the accuracy of cardiac structure segmentation in echocardiography.
• This approach effectively utilizes unlabeled data, aiding in more accurate cardiovascular disease (CVD) diagnosis and screening.
• The method offers a promising solution for improving echocardiographic analysis.