Skip to content

Automated Analysis of In Vivo Confocal Microscopy Corneal Images Using Deep Learning

Brandmark1 Cropped

We are pleased to report that our abstract, Automated Analysis of In Vivo Confocal Microscopy Corneal Images Using Deep Learning, was accepted for presentation at the 2018 Annual Meeting of the Association for Research in Vision and Ophthalmology (ARVO).   The research was performed in collaboration with Professor Joseph Mankowski’s group at Johns Hopkins School of Medicine’s Department of Molecular and Comparative Pathobiology, and Professor Charles McGhee’s group at the University of Auckland, New Zealand.

Quantitative measurements of nerve fibers yield important biomarkers for screening and management of sensory neuropathy in diseases such as diabetes mellitus, HIV, Parkinson’s, and multiple sclerosis. The cornea has the densest set of nerves in the entire body making it an ideal candidate for such assessment.  These measurements, however, are time consuming and subjective when done manually, so an automated method is much preferred.  We extended previous work on automated assessment of in vivo confocal microscopy images of the cornea by replacing our nerve fiber detection algorithm with a deep learning approach.  This has improved overall performance significantly. It has, in particular, improved results in macaque images, a challenging dataset where existing algorithms showed low correlation to manual graders.  This improved robustness, sensitivity, and specificity will drive future studies investigating these parameters as biomarkers for detection and management of a number of disease conditions and, eventually, clinical adoption.