Image denoising for fluorescence microscopy by supervised to self-supervised transfer learning

When using fluorescent microscopy to study cellular dynamics, trade-offs typically have to be made between light exposure and quality of recorded image to balance the phototoxicity and image signal-to-noise ratio. Image denoising is an important tool for retrieving information from dim cell images. Recently, deep learning based image denoising is becoming the leading method because of its promising denoising performance, achieved by leveraging available prior knowledge about the noise model and samples at hand. We demonstrate that incorporating temporal information in the model can further improve the results. However, the practical application of this method has seen challenges because of the requirement of large, task-specific training datasets. In this work, we addressed this challenge by combining self-supervised learning with transfer learning, which eliminated the demand of task-matched training data while maintaining denoising performance. We demonstrate its application in fluorescent imaging of different subcellular structures.