Consensus-guided evaluation of self-supervised learning in echocardiographic segmentation.

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Segmentation

Background:

• Left ventricle segmentation is crucial for cardiac function assessment in echocardiography.
• Deep learning requires large annotated datasets, which are costly and time-consuming to create.
• Self-supervised learning (SSL) offers a way to utilize unlabelled data, but its application in left ventricle segmentation needs further exploration.

Purpose of the Study:

• To investigate the effectiveness of various SSL pretext tasks for echocardiographic segmentation.
• To analyze the impact of unlabelled dataset size and distribution on model pre-training.
• To introduce a novel multi-expert annotated dataset for robust segmentation evaluation.

Main Methods:

• Comparison of different SSL methods for echocardiographic segmentation.
• Evaluation of pre-training performance with varying amounts of unlabelled data.
• Development of a consensus-based annotation strategy using multiple experts to reduce noise.

Main Results:

• Contrastive learning demonstrated superior performance among SSL methods, especially in low-label scenarios.
• AI models pre-trained with SSL and fine-tuned on 15% labelled data showed better agreement with multi-expert consensus than individual experts.
• Excessive unlabelled data during pre-training can negatively impact performance due to redundancy.

Conclusions:

• AI models trained with SSL can generalize effectively across diverse expert annotations.
• This approach leads to more reliable and reproducible cardiac function assessments.
• SSL is a viable strategy to reduce reliance on large, manually annotated datasets for echocardiographic segmentation.