SPARSE data, rich results: Few-shot semi-supervised learning via class-conditioned image translation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning in medical imaging is hindered by limited labeled data.
• Existing methods struggle in low-data regimes, increasing annotation costs.
• Developing effective models with scarce annotations is a critical challenge.

Purpose of the Study:

• To introduce a novel GAN-based semi-supervised learning framework for medical imaging.
• To address the challenge of insufficient labeled training data in low-data regimes.
• To enable robust classification performance with minimal labeled samples.

Main Methods:

• A three-phase training framework integrating a generator, discriminator, and classifier.
• Alternating supervised training on limited labeled data with unsupervised learning via image-to-image translation.
• Ensemble-based pseudo-labeling with confidence weighting and temporal consistency using exponential moving averaging.

Main Results:

• Statistically significant improvements over six state-of-the-art GAN-based semi-supervised methods across eleven MedMNIST datasets.
• Exceptional performance in the extreme 5-shot setting, demonstrating effectiveness with minimal labeled data.
• Consistent superiority across all evaluated settings (5, 10, 20, and 50 shots per class).

Conclusions:

• The proposed framework offers a practical solution for medical imaging applications with prohibitive annotation costs.
• Enables robust classification performance even with extremely limited labeled data.
• Demonstrates the potential of GAN-based semi-supervised learning in data-scarce medical imaging scenarios.