Using Machine Learning and Optical Microscopy Image Analysis of Immunosensors Made on Plasmonic Substrates:...

In this article, we introduce a diagnostic platform comprising an optical microscopy image analysis system coupled with machine learning. Its efficacy is demonstrated in detecting SARS-CoV-2 virus particles at concentrations as low as 1 PFU (plaque-forming unit) per milliliter by processing images from an immunosensor on a plasmonic substrate. This high performance was achieved by classifying images with the support vector machine (SVM) algorithm and the MobileNetV3_small convolutional neural network (CNN) model, which attained an accuracy of 91.6% and a specificity denoted by an F1 score of 96.9% for the negative class. Notably, this approach enabled the detection of SARS-CoV-2 concentrations 1000 times lower than the limit of detection achieved with localized surface plasmon resonance (LSPR) sensing using the same immunosensors. It is also significant that a binary classification between control and positive classes using the MobileNetV3_small model and the random forest algorithm achieved an accuracy of 96.5% for SARS-CoV-2 concentrations down to 1 PFU/mL. At such low concentrations, straightforward screening of newly infected patients may be feasible. In supporting experiments, we verified that texture was the main contributor to the distinguishability of images taken at different SARS-CoV-2 concentrations, indicating that the combination of ML and image analysis may be applied to any biosensor whose detection mechanism is based on adsorption.