Smart Nanoscopy: A Review of Computational Approaches to Achieve Super-Resolved Optical Microscopy

Abstract

The field of optical nanoscopy, a paradigm referring to the recent cutting-edge developments aimed at surpassing the widely acknowledged 200nm-diffraction limit in traditional optical microscopy, has gained recent prominence & traction in the 21 st century. Numerous optical implementations allowing for a new frontier in traditional confocal laser scanning fluorescence microscopy to be explored (termed super-resolution fluorescence microscopy) have been realized through the development of techniques such as stimulated emission and depletion (STED) microscopy, photoactivated localization microscopy (PALM) and stochastic optical reconstruction microscopy (STORM), amongst others. Nonetheless, it would be apt to mention at this juncture that optical nanoscopy has been explored since the mid-late 20 th century, through several computational techniques such as deblurring and deconvolution algorithms. In this review, we take a step back in the field, evaluating the various in silico methods used to achieve optical nanoscopy today, ranging from traditional deconvolution algorithms (such as the Nearest Neighbors algorithm) to the latest developments in the field of computational nanoscopy, founded on artificial intelligence (AI). An insight is provided into some of the commercial applications of AI-based super-resolution imaging, prior to delving into the potentially promising future implications of computational nanoscopy. This is facilitated by recent advancements in the field of AI, deep learning (DL) and convolutional neural network (CNN) architectures, coupled with the growing size of data sources and rapid improvements in computing hardware, such as multi-core CPUs & GPUs, low-latency RAM and hard-drive capacities.

Incorporating the image formation process into deep learning improves network performance in deconvolution applications

“…After fixing, clearing, and sectioning the brain, we located and imaged a region with dense neurite labelling. The size of the brain volume after dual-view reconstruction spanned 5432 × 8816 × 1886 voxels (∼1.4 × 2.3 × 0.5 mm 3 , 168.2 GB in 16-bit format). From this dataset we randomly selected 40 subvolumes, each 256 × 256 × 256 voxels, pairing single-views (input) with dual-view joint deconvolutions (ground truth) to train the network, then applied the trained model to the entire dataset ( Fig.…”

“…All fluorescence images are contaminated by blurring and noise, but this degradation can be ameliorated with deconvolution 1, 2, 3 . For example, iterative Richardson-Lucy deconvolution (RLD) 4, 5 is commonly used in fluorescence microscopy, and is appropriate if the dominant noise source is described by a Poisson distribution.…”

Automated Landslide-Risk Prediction Using Web GIS and Machine Learning Models

“…These learning network-based approaches consist of two components which are made up of a learning network and a result. For the first component, determining the optimal learning network architecture and parameters of all models for the problem is critical as it affects the model performance [ 56 ]. The network is capable of learning linearity or non-linearity data.…”

Highly Efficient and Scalable Framework for High-Speed Super-Resolution Microscopy

“…The efficiency of multi-machines and a GPU approach was also proved in [22]. In addition, a review of computational approaches from the past to the future to obtain super-resolution microscopy is presented in [23].…”