Active learning for left ventricle segmentation in echocardiography.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Cardiology

Background:

• Deep learning for medical image segmentation demands extensive annotated datasets.
• Creating these datasets is costly and time-consuming.
• Active learning offers a solution by selecting informative samples.

Purpose of the Study:

• Investigate active learning for efficient left ventricle segmentation in echocardiography.
• Reduce the burden of sparse expert annotations.
• Introduce and evaluate a novel sampling strategy.

Main Methods:

• Adapt and evaluate various active learning sampling techniques.
• Introduce Optimised Representativeness Sampling (ORS) combining outlier and representative samples.
• Apply methods to echocardiography datasets with sparse annotations.

Main Results:

• Achieved 99% performance using only 20% of labelled data, reducing annotations by 1680 images.
• ORS demonstrated a 70% reduction in annotation effort on a public dataset, outperforming baseline active learning (50% reduction).
• Highlighted dataset-specific performance variations among sampling strategies.

Conclusions:

• Developed a cost-effective active learning approach for echocardiography segmentation.
• Publicly released a novel dataset to advance research in efficient medical image annotation.
• Contributed to understanding efficient annotation strategies in medical image segmentation.