基于无监督深度学习的荧光显微镜图像超分辨率重建  

Fluorescence microscopy images super-resolution reconstruction based on unsupervised deep learning

作　　者：陈东昱 程昱[1] CHEN Dongyu;CHENG Yu(Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]广东工业大学信息工程学院,广州510006

出　　处：《激光杂志》2025年第1期165-172,共8页Laser Journal

摘　　要：针对现有荧光显微镜高质量成对数据集获取难度高、成本高以及低荧光水平下图像超分辨率易产生伪影等问题,提出了一种基于无监督方法的荧光显微镜图像超分辨率重建方法,通过将本文创新的稀疏性提取模块与注意力门融入编码器-解码器网络模型中,实现免高分辨率真值图像且免预先预训练地进行图像超分辨率,并获得更符合人类感官的重建。在公开数据集BioSR与SR-CAO-2上进行了实验,对本方法方法进行了评价。以BioSRs中的Microtubules第一个信噪比等级数据集为例,该方法与经典无监督算法DIP相比,在PSNR上提升了8 dB,在SSIM上提升了0.26,在NIQE上降低了2.5。总体而言,本方法更适合于荧光显微镜的超分辨率任务。Aiming at the problems of high difficulty and high cost in acquiring existing high-quality paired datasets for fluorescence microscopy and the artifacts easily generated by image super-resolution at low fluorescence levels,this paper proposes an unsupervised method based on super-resolution reconstruction of fluorescence microscopy images,which realizes pre-training-free image super-resolution by incorporating this paper's innovative sparsity extraction module and attention gates into encoder-decoder network model,and obtains reconstruction more consistent with human senses.by incorporating the innovative sparsity extraction module and attention gate into the encoder-decoder network model in this paper,to realize pre-training-free image super-resolution reconstruction and to obtain reconstruction more in line with human senses.Experiments are conducted on three different cellular structure datasets from the publicly available dataset BioSR to evaluate the method approach of this paper.Taking the first signal-to-noise level dataset of Microtubules as an example,the method improves 7 dB on PSNR,0.21 on SSIM and reduces 4.6 on NIQE compared to the classical unsupervised algorithm DIP.Overall,the present method is more suitable for super-resolution tasks in fluorescence microscopy.

关 键 词：深度学习 无监督学习 荧光显微镜 深度图像先验 

分 类 号：TN249[电子电信—物理电子学]