Dataset Distillation in Medical Imaging: A Feasibility Study

Area of Science:

• Medical Image Analysis
• Computer Science
• Data Science

Background:

• Data sharing in medical imaging is crucial for model training but faces efficiency challenges.
• Current methods often require transferring entire datasets, limiting collaboration.
• Data distillation, a computer science technique, shows potential for efficient data sharing.

Purpose of the Study:

• To investigate the applicability and effectiveness of data distillation methods in medical imaging.
• To assess the impact of data distillation across diverse medical datasets.
• To identify predictors for successful data distillation performance in this domain.

Main Methods:

• Conducted extensive experiments using various leading data distillation techniques.
• Evaluated methods across multiple medical imaging datasets with varying degrees of data variation.
• Assessed model performance and identified indicators for distillation success.

Main Results:

• Data distillation significantly reduces medical dataset size while preserving model performance.
• Comparable results were achieved compared to using the full dataset.
• A small, representative image sample can reliably indicate successful distillation.

Conclusions:

• Data distillation is a viable and effective method for efficient and secure medical data sharing.
• This approach can facilitate enhanced collaborative research and clinical applications.
• The findings suggest potential for optimized medical data management and analysis.