Dataset Distillation in Medical Imaging: A Feasibility Study

TL;DR

This study examines the feasibility of dataset distillation for medical imaging to enable efficient, privacy-preserving data sharing. It evaluates two distillation methods, DC and MTT, across multiple modalities and eight MedMNIST datasets, using both a distillation-simulation setup and per-dataset condensation. The results show that DC often achieves near full-dataset performance with substantially smaller synthetic sets, while MTT is generally less robust; a strong correlation between distilled-set and random-subset performance provides a practical predictor for distillation success. Overall, the work demonstrates the potential of data distillation to facilitate secure, collaborative medical research and highlights areas for method-specific improvements in medical contexts.

Abstract

Data sharing in the medical image analysis field has potential yet remains underappreciated. The aim is often to share datasets efficiently with other sites to train models effectively. One possible solution is to avoid transferring the entire dataset while still achieving similar model performance. Recent progress in data distillation within computer science offers promising prospects for sharing medical data efficiently without significantly compromising model effectiveness. However, it remains uncertain whether these methods would be applicable to medical imaging, since medical and natural images are distinct fields. Moreover, it is intriguing to consider what level of performance could be achieved with these methods. To answer these questions, we conduct investigations on a variety of leading data distillation methods, in different contexts of medical imaging. We evaluate the feasibility of these methods with extensive experiments in two aspects: 1) Assess the impact of data distillation across multiple datasets characterized by minor or great variations. 2) Explore the indicator to predict the distillation performance. Our extensive experiments across multiple medical datasets reveal that data distillation can significantly reduce dataset size while maintaining comparable model performance to that achieved with the full dataset, suggesting that a small, representative sample of images can serve as a reliable indicator of distillation success. This study demonstrates that data distillation is a viable method for efficient and secure medical data sharing, with the potential to facilitate enhanced collaborative research and clinical applications.

Figures (5)

• Figure 1: Difference between natural and medical datasets
• Figure 2: General working pipeline for this paper. We investigate the effectiveness of distillation on medical datasets by designing multiple experiments to answer these questions.
• Figure 3: Distillation simulation on a large inter-class variance integrated medical dataset. The original datasets, from left to right, are TissueMNIST, PathMNIST, OrganSMNIST, OrganCMNIST, OrganAMNIST, OCTMNIST, DermaMNIST, and BloodMNIST, each treated as a single class in this integrated dataset. The right figure shows that higher distillation accuracy highlights the potential for increasing inter-class variance in specific medical datasets for improved distillation.
• Figure 4: We test distillation methods on 9 different medical datasets, and most of large variation medical dataset shows relatively ideal distillation performance as is shown like a), and most of small variation datasets, such as b), shows lower effectiveness in distillation.
• Figure 5: This figure shows the correlation between MTT/DC and Random Selected method. Comparing MTT and DC's performance on IPC$=$50 referred to randomly selected 50 images from original datasets, MTT shows higher correlation than DC, which indicates its possible better distillation performance on higher IPC. This can answer the third question in our main pipeline.