Compound Figure Separation of Biomedical Images: Mining Large Datasets for Self-supervised Learning

Area of Science:

• Medical image analysis
• Artificial intelligence
• Computer vision

Background:

• Self-supervised learning (SSL) requires large-scale unannotated images for generalizable AI models in medical imaging.
• Collecting such datasets is challenging for individual research labs.
• Online resources offer vast image collections, but medical publications often contain compound figures with subplots.

Purpose of the Study:

• To develop a framework for separating compound figures into individual images for downstream AI learning.
• To reduce the reliance on extensive bounding box annotations for training.
• To evaluate the effectiveness of SSL combined with compound figure separation in medical image analysis.

Main Methods:

• Proposed SimCFS (Simple Compound Figure Separation) framework.
• Introduced a simulation-based training approach to minimize annotation needs.
• Developed a novel side loss function optimized for figure separation.
• Implemented an intra-class image augmentation technique to simulate challenging cases.

Main Results:

• SimCFS achieved state-of-the-art performance on the ImageCLEF 2016 Compound Figure Separation Database.
• Self-supervised learning models pretrained with separated figures improved downstream image classification accuracy.
• Contrastive learning algorithms benefited from the enhanced dataset.

Conclusions:

• SimCFS effectively separates compound figures, enabling better utilization of online medical images for AI training.
• The approach significantly reduces annotation burden while improving model performance.
• This study demonstrates the first successful integration of SSL with compound image separation for medical AI.