TagGAN: A generative model for data tagging.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Computer Vision

Background:

• Accurate pixel-level disease identification is crucial for diagnosis and monitoring.
• Conventional AI methods struggle with limited pixel-level annotations and lack transparency.
• Existing techniques require binary masks, which are often unavailable.

Purpose of the Study:

• To develop a weakly-supervised framework for fine-grained disease map generation using only image-level labels.
• To enhance the interpretability of diagnostic AI through precise disease lesion visualization.
• To automate binary mask generation for radiologist assistance.

Main Methods:

• Proposed TagGAN, a Generative Adversarial Networks (GANs)-based framework for weakly-supervised disease map generation.
• Employed domain translation to generate pixel-level disease maps from abnormal to normal image representations.
• Subtracted generated disease maps from abnormal images to create normal counterparts, preserving anatomical details.

Main Results:

• TagGAN successfully generated fine-grained disease maps without requiring pixel-level annotations.
• The framework demonstrated enhanced interpretability by visualizing disease-specific regions.
• Achieved state-of-the-art performance, outperforming existing methods by over 6% in identifying disease-specific pixels on benchmark datasets (CheXpert, TBX11K, COVID-19).

Conclusions:

• TagGAN offers a robust solution for weakly-supervised disease map generation in medical imaging.
• The model significantly improves the accuracy of disease localization and enhances AI transparency.
• TagGAN reduces radiologist workload by eliminating the need for manual binary mask annotation during training.