3D-B2U: Self-supervised Fluorescent Image Sequences Denoising

Fluorescence imaging can reveal the spatiotemporal dynamics of life activities. However, fluorescence image data suffers from photon shot noise due to a limited photon budget. Therefore, denoising fluorescence image sequences is an important task. Existing self-supervised methods solve the problem of complex parameter tuning of non-learning methods and the problem of requiring a large number of noisy-clean image pairs for supervised learning and become state-of-the-art methods for fluorescent image sequences denoising. However, they aim at 2D data, which cannot make good use of the increased time dimension information of fluorescence data compared with single image data. Besides, they still use paired noisy data to train models, and the strong prior information brought by paired data may lead to the overfitting of the model. In this work, we extend existing self-supervised methods to 3D and propose a 3D global masker that introduces a visible blind-spot structure based on 3D convolutions to avoid identity mapping while fully utilizing the input data information. Our method makes reasonable use of time dimension information and enables the task of self-supervised denoising on fluorescent images to mine information from the input data itself. Experimental results show that our method achieves a better denoising effect for fluorescent image sequences.