Deep Learning-based Segmentation and Quantification of Nerve Fibers and Dendritic cells in Confocal Microscopy of the Cornea

Abstract

Purpose : In vivo confocal microscopy (IVCM) is a non-invasive optical imaging modality that enables a histological visualization and provides morphometric information about corneal nerve fiber (CNF) and dendritic cell (DC), that aid clinicians to diagnosis of inflammatory corneal diseases. Quantification of CNF and DCs requires manual annotation or semi-automatic approaches, which are time consuming, non-reproducible, and laborious. The purpose of this research study was to develop deep learning-based models to segment and quantify CNF and DC in IVCM automatically, therefore reducing inter-or intra-observer variability and time associated with manual perception to analyze larger volumes of clinical images.

Methods : A CNF segmentation and DC detection model were developed based on deep learning algorithms U-Net and Mask R-CNN respectively. The CNF segmentation model was trained with 1036 and tested on 183 images while the DC detection model was trained with 446 and tested on 50 images. To analyze CNF morphology, number of nerves, number of branching points, nerves density, nerves length, and tortuosity were measured. An automatic Python-based software was written to compute the morphometric parameters directly from the binary segmented image produced by deep learning model. Moreover, Bland-Altman's statistical analysis was performed to determine the consistency between automatic segmentation and manual annotation.

Results : The CNF segmentation model reliably segments the testing images with an average 83% sensitivity and 91% specificity while the DC detection model detects DCs with an average 92% precision, 95% recall, and 93% F1 score. In Bland-Altman’s analysis, the mean of automatic and manual segmentation for all morphometric parameters was close to 0, and more than 95% of the values were within the limit of agreement. To segment and morphometric evaluation of CNF, our developed software took on average 4.5 seconds per image while to detect and count the DCs took on average 3 seconds.

Conclusions : Our developed deep learning-based models demonstrated high consistency between automatic and manual segmentation of IVCM images with rapid speed. The results show that the system has the potential to be implemented into clinical practice for CNF segmentation and DC detection in IVCM.

This is a 2021 ARVO Annual Meeting abstract.