Self-supervised learning for label-free segmentation in cardiac ultrasound

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiac chamber segmentation from ultrasound is crucial but challenging due to manual labor and reproducibility issues.
• Supervised deep learning methods require manual annotations, which are time-consuming and error-prone.

Purpose of the Study:

• To develop a self-supervised segmentation pipeline for cardiac ultrasound analysis.
• To enable accurate and reproducible measurements of cardiac chambers without manual labels.
• To validate the clinical utility and scalability of the developed method.

Main Methods:

• A self-supervised segmentation pipeline combining computer vision, clinical knowledge, and deep learning was developed.
• The pipeline was trained on 450 echocardiograms and validated on 18,423 echocardiograms, including external datasets.
• Performance was evaluated by comparing pipeline-derived measurements to clinical measurements and cardiac MRI (gold standard).

Main Results:

• The pipeline achieved high accuracy in segmenting cardiac chambers, with an average Dice score of 0.89 for the left ventricle.
• Measurement prediction accuracy (r²=0.55-0.84) was comparable to inter-clinician variability and supervised learning.
• Correlation between pipeline and MRI measurements was similar to that of clinical echocardiograms.

Conclusions:

• The developed self-supervised pipeline offers a clinically valid, manual-label-free, and scalable solution for cardiac ultrasound segmentation and measurement.
• This approach significantly reduces the labor and improves the reproducibility of cardiac chamber analysis.
• The method holds promise for widespread clinical adoption and advancement in echocardiography.