Self-supervised Assisted Active Learning for Skin Lesion Segmentation.

Area of Science:

• Biomedical image segmentation
• Machine learning in medical imaging
• Active learning strategies

Background:

• Label scarcity is a significant challenge in biomedical image segmentation, increasing annotation costs and time.
• Active learning (AL) aims to reduce these costs by selecting informative samples for annotation.
• Current AL methods often rely on random initialization, leading to redundant samples and higher costs.

Purpose of the Study:

• To address the limitations of random initialization in active learning for biomedical image segmentation.
• To propose a novel self-supervised assisted active learning framework for the cold-start setting.
• To reduce annotation costs and improve efficiency in medical image analysis.

Main Methods:

• A self-supervised learning (SSL) approach is used to pre-train the segmentation model.
• Latent features from SSL are employed for sample selection through clustering, without requiring initial labels.
• The framework is evaluated on the skin lesion segmentation task.

Main Results:

• The proposed self-supervised assisted active learning framework achieves promising performance.
• Substantial improvements are demonstrated compared to existing baseline methods.
• The method effectively reduces annotation costs by smart sample selection in the cold-start setting.

Conclusions:

• The novel framework successfully overcomes the challenges of label scarcity and high annotation costs in biomedical image segmentation.
• Self-supervised learning pre-training combined with latent feature clustering offers an efficient approach for active learning.
• This method holds significant clinical relevance by saving annotation resources in practice.