MID3A: Microscopy Image Denoising meets Differentiable Data Augmentation

Image restoration is an important task in a wide variety of topics. Especially in the domain of microscopy images, several content-aware image restoration (CARE) methods have arisen to improve the interpretability of acquired data. One of the main problems is the presence of high levels of noise that must be removed before any further post-processing or analysis happens. To solve this issue, we propose a simple and yet effective framework consisting of a generative adversarial network (GAN) coupled with a regularization term (differentiable data augmentation) that highly increases the quality of the denoising for three different well-known microscopy imaging data sets. We also introduce two structure preserving loss terms (Structural Similarity Index and Total Variation loss) that, added to our framework, help to further improve the quality of the results. In addition, we prove that our method is able to generalize well when trained on one dataset and used in another. Finally, we show that we can drastically reduce the amount of training data while retaining the quality of the denoising, thus alleviating the burden of acquiring paired data and enabling few-shot learning.