DEEP NEURAL NETWORKS FOR IMAGE SUPER-RESOLUTION IN OPTICAL MICROSCOPY BY USING MODIFIED HYBRID TASK CASCADE U-NET

Abstract

Due to the optical diffraction limit, the resolution of a wide-field (WF) microscope cannot easily go below a few hundred nanometers. Super-resolution microscopy has the disadvantages of high cost, complex optical equipment, and high experimental environment requirements. Deep-learningbased super-resolution (DLSR) has the advantages of simple operation and low cost, and has attracted much attention recently. Here we propose a novel DLSR model named Modified Hybrid Task Cascade U-Net (MHTCUN) for image super-resolution in optical microscopy using the public biological image dataset BioSR. The MHTCUN has three stages, and we introduce a novel module named Feature Refinement Module (FRM) to extract deeper features in each stage. In each FRM, a U-Net is introduced to refine the features, and the Fourier Channel Attention Block (FCAB) is introduced in the U-Net to learn the high-level representation of the high-frequency information of different feature maps. Compared with six state-of-the-art DLSR models used for single-image super-resolution (SISR), our MHTCUN achieves the highest signal-to-noise ratio (PSNR) of 26.87 and structural similarity (SSIM) of 0.746, demonstrating that our MHTCUN has achieved the state-of-the-art in DLSR. Compared with the DLSR model DFCAN used for image super-resolution in optical microscopy specially, MHTCUN has a significant improvement in PSNR and a slight improvement in SSIM on BioSR. Finally, we finetune the trained MHTCUN on the other biological images. MHTCUN also shows good performance on denoising, contrast enhancement, and resolution enhancement.