Segmenting Cardiac Ultrasound Videos Using Self-Supervised Learning.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Cardiovascular diagnostics

Background:

• Deep learning models for cardiac ultrasound segmentation often require extensive labeled data, limiting their generalizability and performance.
• Insufficient data volume can lead to poor generalization across different equipment, clinics, and clinicians.

Purpose of the Study:

• To improve cardiac ultrasound segmentation models using unlabeled data and sparse labels.
• To reduce the reliance on large amounts of dense, pixel-level annotations for training.
• To enhance the accuracy of cardiac Ejection Fraction (EF) estimation.

Main Methods:

• Utilized self-supervised learning on unlabeled data to learn recurrent anatomical representations.
• Employed supervised local contrastive learning on sparse labels to improve segmentation.
• Implemented supervised fine-tuning for segmenting temporal anatomical features to estimate EF.
• Validated performance using DeepLabv3+ and Attention U-Net models.

Main Results:

• Pretraining network weights with self-supervised learning followed by supervised contrastive learning significantly outperformed training from scratch.
• The proposed methods improved cardiac ejection fraction evaluation compared to existing techniques.
• Demonstrated enhanced segmentation performance and reduced annotation burden.

Conclusions:

• Self-supervised and contrastive learning strategies effectively improve cardiac ultrasound segmentation and EF estimation.
• This approach offers a viable solution for developing robust cardiac function evaluation tools with reduced data requirements.
• The findings have direct clinical relevance for assisting physicians in cardiac function assessments.