Fluorescence Image Denoising Based on Self-supervised Deep Learning

Fluorescence microscopy imaging technology is a crucial imaging technology widely used in biomedical fields such as brain science. However, its images often have random noise without a fixed pattern, causing a series of problems such as image degradation, image segmentation, and classification misalignment. In order to obtain higher signal-to-noise ratio imaging, the input of higher doses of photons increases the cost and causes adverse effects such as photobleaching, phototoxicity, or tissue heating. Therefore, researchers have studied many traditional denoising methods and deep learning-based denoising methods. Traditional denoising methods have the advantage of fast and straightforward processing and tend to perform well when the noise level of the image is low, or the noise distribution is known. However, its effectiveness and generalization are minimal when the noise level is high and randomly distributed without a fixed modality. On the contrary, image denoising methods based on self-supervised deep learning can extract high-dimensional features of images, have better effectiveness and robustness in practical denoising applications. Morever, these methods reduce the need for manual annotation of large-scale data, making timely research on cutting- edge pathology possible when there are insufficient labeled data. In this paper, to tackle the denoising problem of fluorescence microscopy images, two classic traditional denoising methods, Wiener filtering, and wavelet transformation, and methods based on deep learning, DeepCAD, are used for comparative experiments. The experimental results show that the deep learning method has good effectiveness and generalization in dealing with high-level random noise, and it performs better than traditional denoising methods.