Privacy dataset distillation for medical imaging using class centralization and covariance matching for OCTMNIST

In this study, we introduce a novel dataset distillation method applied to medical imaging, specifically using the OCTMNIST dataset. Our approach incorporates advanced techniques such as class centralization and covariance matching constraints to efficiently synthesize smaller, yet representative, datasets from a larger corpus of optical coherence tomography images. The efficacy of our method is quantitatively validated against standard distillation techniques such as DC, DM, and DSA, showcasing superior performance in terms of classification accuracy and model stability. Additionally, we emphasize privacy protection by employing metrics such as the L2 norm and Structural Similarity Index Measure (SSIM) to ensure that the distilled images do not retain sensitive information. The results illustrate our method's ability to maintain high accuracy with fewer images while significantly enhancing data privacy, making it ideal for medical imaging applications where data sensitivity is paramount.