Privacy preserving data distillation in medical imaging with multidimensional matching on PATHMNIST

In the domain of medical image analysis, the privacy of patient data is paramount, yet the need for extensive datasets to train robust models is ever-increasing. This paper introduces a novel dataset distillation approach that leverages multidimensional matching methods, including distribution, gradient, and trajectory matching, to generate synthetic datasets that preserve the utility of original medical datasets while enhancing privacy protection. Applied to the PATHMNIST dataset, a colon pathology benchmark, our method not only achieves superior model performance compared to existing dataset distillation techniques but also significantly improves the privacy of the distilled images, as evidenced by lower L2 norms in pixel-level comparisons. Our findings demonstrate that our approach can serve as a robust framework for generating training-ready datasets that adhere to privacy constraints inherent in medical data applications.