Facile Quantification of Nanosized Bioparticles in Bright-field Micrograph via Deep Learning

Abstract

The pathogen is considered a critical threat to public health, especially immunocompromisedpeople such as babies, and the elderly. Initiating the apt therapy to the defected patients before turning into thepeak viral load is significant to prevent their fatal damage. However, most point-of-care diagnostics areunsuitable for counting the concentration of the virus [1]. Accordingly, various bioimaging techniques aresuggested to visualize and/or count biological samples like histopathology [2]. Despite their superb resolutionand precise quantitation, the need for sophisticated equipment or professional manpower stunts the broadinterest.Explosive interest in artificial intelligence (AI) induces profuse interdisciplinary studies, especially in biomedicalengineering and clinical medicine recently [3]. Due to the property of pure convolution and superior objectcategorization skills, convolutional neural networks (CNNs) that model human vision has been demonstrated incomprehensive clinical data to predict patient diagnoses. Even though CNNs produce high capabilities for theassigned task, plenty of pre-processing and the requirement of expert interpretation are still mandated.Tri-layer resonator is adopted as the optical solution to break through weak light-matter interaction in far-fieldoptics. By harnessing thin-film interference, the modulated light produces a strong resonance in thebioparticles of low refractive index and nanoscale size and leads to the perception of targeted analytes.Additionally, the aggregation of biofunctionalized nanoparticles in drop-casted solution due to hydrodynamicsof an evaporating droplet (e.g., the Marangoni flow) synergizes with the slow-light effect so that the invisibleanalytes can be discovered with definite chromatic information even through the usual optical microscope(OM). During the nanofabrication and surface-functionalizing, however, the residues and defects randomlyexist on the surface of the designed biosensor. Hence, we introduce the CNN that establishes a correlationbetween the chromatic information of optical micrographs and the ground-truth of corresponding scanningelectron microscope (SEM) images.Here, we showcase the chromatic immunoassay system consisting of the biosensor that is optically optimizedfor general zoonotic virus and deep learning that is robustly constructed with 1596 pairs of OM images andmatching SEM images. The hard negative samples of the fabrication fluctuation including the impurities andthe defects are trained to identify the desired particles automatically, thus averting the false-positive andenabling accurate quantitative analysis. The image-driven biosensing system validates the limit of viraldetection of 104 copies/mL which is lower than the rapid antigen test using lateral flow immunoassay (LFIA)and achieves high sensitivity and specificity for the diverse bioparticles modeling Zika virus, Monkeypox virus,and Mumps virus. The presented immunoassay platform may also be amenable to microscopic hazardousfragments such as metal oxide nanoparticles and microplastics.In this study, we have shown that CNN-based bright-field micrograph analysis allows not only intuitiveimmunoassay but also quantification of minuscule subjects. Nanosized bioparticles are detected with astraightforward antibody-antigen reaction, which does not require either labeling or amplification. A single dropof the solution enables visceral recognition through the vision. Owing to the purity of the CNN, the variousanalytes whose diameters are below the diffraction limit can be estimated with the little training set. We alsohave confirmed that the presented immunoassay quantifies the concentration accurately within the averageviral load from SARS-CoV-2-infected patients. We believe that the streamlined quantitative system isapplicable to numerous nanoscale dicey particles.