SimLVSeg: Simplifying Left Ventricular Segmentation in 2-D+Time Echocardiograms With Self- and Weakly Supervised...

Area of Science:

• Medical imaging
• Artificial intelligence in healthcare
• Cardiovascular imaging

Background:

• Automatic left ventricle (LV) segmentation in echocardiograms is hindered by limited annotation data.
• Clinicians typically annotate only a few frames, posing a challenge for deep learning models.

Purpose of the Study:

• To introduce SimLVSeg, a novel paradigm for consistent LV segmentation in echocardiogram videos using sparsely annotated data.
• To enable video-based networks for improved LV segmentation accuracy and efficiency.

Main Methods:

• SimLVSeg employs a two-stage training process: self-supervised pre-training with temporal masking and weakly supervised learning.
• The self-supervised stage leverages cyclic patterns in unannotated echocardiogram frames.
• The weakly supervised stage refines segmentation using sparse clinical annotations.

Main Results:

• SimLVSeg achieved a 93.32% Dice score on the EchoNet-Dynamic dataset, outperforming existing methods.
• The method demonstrated superior efficiency compared to state-of-the-art solutions.
• Out-of-distribution testing on the CAM US dataset confirmed SimLVSeg's generalizability.

Conclusions:

• SimLVSeg offers excellent performance for LV segmentation with reduced computational cost.
• Video-based networks represent a promising direction for reliable echocardiogram segmentation.
• The developed SimLVSeg framework and code are publicly available for research use.