Mu-net: Multi-scale U-net for two-photon microscopy image denoising and restoration. (May 2020)

• Record Type: Journal Article
• Authors: Lee, Sehyung; Negishi, Makiko; Urakubo, Hidetoshi; Kasai, Haruo; Ishii, Shin
• Is Part Of: Neural networks. Volume 125(2020)
• Page Start: 92
• Page End: 103
• Publication Date: 2020-05
• Subjects: Image denoising -- Two-photon microscopy image -- Deep learning -- U-net -- GAN; Neural computers -- Periodicals; Neural networks (Computer science) -- Periodicals; Neural networks (Neurobiology) -- Periodicals; Nervous System -- Periodicals; Ordinateurs neuronaux -- Périodiques; Réseaux neuronaux (Informatique) -- Périodiques; Réseaux neuronaux (Neurobiologie) -- Périodiques; Neural computers; Neural networks (Computer science); Neural networks (Neurobiology); Periodicals; 006.32
• DOI: 10.1016/j.neunet.2020.01.026
• Languages: English
• ISSNs: 0893-6080

Abstract

Abstract: Advances in two two-photon microscopy (2PM) have made three-dimensional (3D) neural imaging of deep cortical regions possible. However, 2PM often suffers from poor image quality because of various noise factors, including blur, white noise, and photo bleaching. In addition, the effectiveness of the existing image processing methods is limited because of the special features of 2PM images such as deeper tissue penetration but higher image noises owing to rapid laser scanning. To address the denoising problems in 2PM 3D images, we present a new algorithm based on deep convolutional neural networks (CNNs). The proposed model consists of multiple U-nets in which an individual U-net removes noises at different scales and then yields a performance improvement based on a coarse-to-fine strategy. Moreover, the constituent CNNs employ fully 3D convolution operations. Such an architecture enables the proposed model to facilitate end-to-end learning without any pre/post processing. Based on the experiments on 2PM image denoising, we observed that our new algorithm demonstrates substantial performance improvements over other baseline methods.