Deep-learning based analysis of in-vivo confocal microscopy-derived metrics of the corneal nerve whorl in eyes with neuropathic ocular pain and controls

Abstract

Purpose : To compare in-vivo confocal microscopy (IVCM) metrics of sub-basal corneal nerves, immune cells, and neuromas patients with neuropathic ocular pain (NOP) and controls using deep learning-based automated segmentation and assess the validity of corneal nerve whorl detection

Methods : In-vivo confocal microscopy images of the corneal whorl were collected from 97 eyes (49 NOP, 48 Controls). Inclusion criteria for neuropathic ocular pain included Ocular Surface Disease Index (OSDI) > 13, minimal to no ocular surface fluorescein staining (less or equal to 1 on NEI scale), and presence of ocular pain.
IVCM images were exclusively collected from the corneal whorl, located in the inferonasal portion of the sub-basal nerve plexus. The images were automatically segmented by a deep learning algorithm trained on 207 manually segmented images from 69 eyes with ocular surface diseases (OSD). The algorithm labeled each pixel as pertaining to nerve, immune cell, neuroma, or background. Using the segmentation, we subsequently calculated the nerve density, average nerve thickness, nerve tortuosity, immune cell count, and neuroma count within an image frame (16000 μm²).

Results : The mean age of the subjects was 60 ± 12.5 (range 38 to 82 years), and 75% of them were female. 60 NOP eyes and 68 control eyes were imaged, yet focused corneal whorl images were obtained in 49 eyes and 48 eyes respectively.

The neuroma count was significantly higher in the NOP group as compared to the control group (2.23± 0.39 neuromas/frame vs 1.26 ± 0.20 neuromas/frame, P=0.03). Average nerve density was also significantly lower in the NOP group as compared to the controls (4616 μm ± 383.9 μm/frame vs 7832 μm ± 424.4 μm/frame, P=0.01).
Immune cell counts, nerve thickness and nerve tortuosity were not statistically different in the NOP group as compared to the control group (P=0.21, P=0.55, P=0.95, respectively)

Conclusions : Using IVCM corneal subbasal nerve plexus imaging with a standard landmark (the whorl) and deep learning- based automated segmentation, neuropathic ocular pain can be distinguished from controls on IVCM by an increased number of neuromas and lower average nerve density.
These findings validate similar observations in standard IVCM imaging and suggest that the corneal whorl can serve as a landmark biomarker for metric evaluation.

This abstract was presented at the 2023 ARVO Annual Meeting, held in New Orleans, LA, April 23-27, 2023.