A Self Supervised StyleGAN for Image Annotation and Classification With Extremely Limited Labels.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Deep learning success relies on large annotated datasets, but labeling is costly.
• Limited annotations degrade deep learning model performance.
• Self-supervised learning (SSL) leverages unlabeled data for feature representation.

Purpose of the Study:

• To propose SS-StyleGAN, a self-supervised framework for image annotation and classification with extremely small labeled datasets.
• To integrate self-supervision into the StyleGAN architecture via an encoder for latent space embedding.

Main Methods:

• Developed SS-StyleGAN by incorporating an encoder into StyleGAN to learn disentangled latent space representations.
• Utilized the learned latent space for intelligent selection of data representatives for labeling.
• Applied the framework to image classification tasks with minimal annotated data.

Main Results:

• Achieved strong classification performance using very small labeled datasets (e.g., 50 and 10 samples).
• Demonstrated the effectiveness of the approach for COVID-19 identification.
• Showcased superiority in liver tumor pathology identification tasks.

Conclusions:

• SS-StyleGAN effectively mitigates the need for extensive labeled data in deep learning.
• The method offers a viable solution for image classification in domains with scarce annotations.
• Validated the approach's efficacy in critical medical imaging applications.