Compressed gastric image generation based on soft-label dataset distillation for medical data sharing

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Data Compression

Background:

• Efficient medical data sharing is crucial for healthcare information exchange and developing accurate computer-aided diagnosis systems.
• Challenges include large dataset sizes, substantial memory requirements for deep convolutional neural network (DCNN) models, and patient privacy concerns.
• These factors hinder effective medical data sharing and collaboration.

Purpose of the Study:

• To propose a novel soft-label dataset distillation method for enhancing medical data sharing.
• To address the inefficiencies and security risks associated with current medical data sharing practices.
• To enable secure and efficient cross-agency healthcare information flow.

Main Methods:

• Distillation of valid information from medical image data to generate compressed images with varied distributions for anonymous sharing.
• Extraction of essential weights from DCNN models to reduce memory footprint for efficient storage and sharing.
• Generation of visually anonymized compressed images that do not contain patient-specific private information.

Main Results:

• Compression of tens of thousands of images into a few soft-label images.
• Reduction of trained DCNN model size to a fraction of its original size.
• Demonstration of high-detection performance using a small number of compressed images.

Conclusions:

• The proposed soft-label dataset distillation method significantly improves the efficiency of medical data sharing.
• The method enhances the security of medical data by anonymizing images and reducing model size.
• Experimental results confirm the effectiveness of the approach in practical applications.