Fast Contactless Detection of Post-COVID-19 from Eye Microvascular Images Using Deep Convolutional Neural Networks

Background/Objectives: Artificial intelligence (AI) and its subfield deep learning (DL) have rapidly evolved into a central tool in modern medicine. The purpose of this work was to examine if DL neural networks can discriminate efficiently the microvessel network of post-COVID-19 patients from healthy individuals from. Methods: A non-contact, digital slit-lamp video capillaroscopy system was used to record high magnification images form the bulbar conjunctival microcirculation of 12 COVID-19 survivors (named “COVID-19 Group”) and 12 healthy volunteers (named “Control Group”). Four pretrained convolutional neural networks (CNNs) were fine-tuned by transfer learning and their performance was assessed by standard binary classification evaluation criteria. Results: A scene-centric CNN named GoogLeNet-Places365 excelled on all evaluation criteria with an average testing accuracy, sensitivity, specificity and AUC (area under the curve) of 92%, 92%, 91%, and 0.971, respectively. Conclusions: Post-COVID effects on the eye microcirculation can be detected by deep CNNs, and there is now evidence for the first time, that AI could provide a risk-free, painless, contactless, fast, and accurate detection method of viral effects that does not depend on the optical clarity of the eye.