Multiple-Image Deep Learning Analysis for Neuropathy Detection in Corneal Nerve Images

Purpose: Automated classification of corneal confocal images from healthy subjects and diabetic subjects with neuropathy. Methods: Over the years, in vivo confocal microscopy has established itself as a rapid and noninvasive method for clinical assessment of the cornea. In particular, images of the subbasal nerve plexus are useful to detect pathological conditions. Currently, clinical information is derived through a manual or semiautomated process that traces corneal nerves and achieves their descriptors (eg, density and tortuosity). This is tedious and subjective. To overcome this limitation, a method based on a convolutional neural network (CNN) for the classification of images from healthy subjects and diabetic subjects with neuropathy is proposed. The CNN simultaneously analyzes 3 nonoverlapping images, from the central region of the cornea. The algorithm automatically extracts features, without the need for neither nerve tracing nor parameter extraction nor montage/mosaicking, and provides an overall classification for each image trio. Results: On a dataset composed by images from 50 healthy subjects and 50 subjects with neuropathy, the algorithm achieves a classification accuracy of 96%. The proposed method improves the results obtained using a traditional method that traces nerves and evaluates their density and tortuosity. Conclusions: The proposed method provides a completely automated analysis of corneal confocal images. Results demonstrate the potentiality of the CNN in identifying clinically useful features for corneal nerves by analysis of multiple images.