Active Image Synthesis for Efficient Labeling.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Deep neural networks show great promise in manufacturing and healthcare.
• Limited data availability and high collection costs hinder AI adoption in these fields.
• Small-data learning tasks require efficient methods to overcome data limitations.

Purpose of the Study:

• To introduce AISEL, an active image synthesis method for efficient labeling.
• To improve the performance of small-data learning tasks using AISEL.
• To generate a complementary dataset with labels acquired via a physics-based method.

Main Methods:

• Utilizing a bidirectional generative invertible network (GIN) for feature extraction and virtual image generation.
• Incorporating physical knowledge into the dataset through physics-based labeling.
• Actively sampling virtual images to explore uncertain regions and the entire image space.
• Demonstrating the interpretability of GIN theoretically and experimentally.

Main Results:

• AISEL generates physically meaningful virtual images and extracts interpretable features.
• The method effectively exploits uncertain regions and explores the image space.
• GIN shows clear visual improvements over existing benchmarks.
• The AISEL framework demonstrates effectiveness in aortic stenosis applications.

Conclusions:

• AISEL significantly reduces labeling costs by 90% in medical applications.
• The method achieves a 15% improvement in prediction accuracy.
• AISEL offers an efficient solution for small-data learning challenges in data-scarce domains.
• The integration of physical knowledge enhances the performance of AI models.