Evaluation of Swin Transformer and Knowledge Transfer for Denoising of Super-resolution Structured Illumination Microscopy Data

Background: Convolutional neural network (CNN)-based methods have shown excellent performance in denoising and reconstruction of super-resolved structured illumination microscopy (SR-SIM) data. Therefore, CNN-based architectures have been the focus of existing studies. However, Swin Transformer, an alternative and recently proposed deep learning-based image restoration architecture, has not been fully investigated for denoising SR-SIM images. Furthermore, it has not been fully explored how well transfer learning strategies work for denoising SR-SIM images with different noise characteristics and recorded cell structures for these different types of deep learning-based methods. Currently, the scarcity of publicly available SR-SIM datasets limits the exploration of the performance and generalization capabilities of deep learning methods.

Results: In this work, we present SwinT-fairSIM, a novel method based on the Swin Transformer for restoring SR-SIM images with a low signal-to-noise ratio. The experimental results show that SwinT-fairSIM outperforms previous CNN-based denoising methods. Furthermore, as a second contribution, two types of transfer learning-namely, direct transfer and fine-tuning-were benchmarked in combination with SwinT-fairSIM and CNN-based methods for denoising SR-SIM data. Direct transfer did not prove to be a viable strategy, but fine-tuning produced results comparable to conventional training from scratch while saving computational time and potentially reducing the amount of training data required. As a third contribution, we publish four datasets of raw SIM images and already reconstructed SR-SIM images. These datasets cover two different types of cell structures, tubulin filaments and vesicle structures. Different noise levels are available for the tubulin filaments.

Conclusion: The SwinT-fairSIM method is well suited for denoising SR-SIM images. By fine-tuning, already trained models can be easily adapted to different noise characteristics and cell structures. Furthermore, the provided datasets are structured in a way that the research community can readily use them for research on denoising, super-resolution, and transfer learning strategies.