Comparison of manual and fully automatic cell hexagonality measure in corneal endothelium images in transplanted corneas post Descemet Stripping Automated Endothelial Keratoplasty (DSAEK)

Abstract

Purpose : We developed deep-learning-based automatic cell segmentation and a hexagonality (HEX) estimation method for analysis of corneal endothelial cell (EC) images taken 6 months to 3 years post DSAEK. As manual analysis is very laborious, successful automatic estimation of HEX would allow comprehensive computational analysis of large datasets of over 1000s of EC images.

Methods : We used a subset of 130 specular microscopy images from 26 eyes collected from the Cornea Preservation Time Study (CPTS). Trained readers performed traditional corner analysis using the CAS/EB software (HAI Labs Inc., Lexington, MA) to generate cell border contours by marking the corners of individual cells. The percentage of hexagonal cells (HEX) output from the software was recorded for each annotated image. In order to automatically compute the HEX metric for a given image, we first trained a deep neural network algorithm (U-Net) to segment ECs in the microscopy images. 65 images were used as the held-out test set. Next, segmentations from the network were binarized and cleaned using morphological operations. Finally, results of connected components analysis on the predicted masks were fed as inputs to a shape detection algorithm that estimates the percentage of hexagonal cells.

Results : From the 65 test images, we found no statistically significant difference between the HEX measure from our fully automatic deep learning based segmentation approach and the ground truth reader annotations (p = 0.90). Over our heterogeneous test data, HEX values were 56.77 ± 0.91 and 56.91 ± 1.11 for manual and automated, respectively.

Conclusions : We found that our shape detection algorithm, based on an iterative end point fit method, can closely estimate cell HEX in an automatic fashion on an initial discovery dataset containing 65 clinical quality specular microscopy images. Further work is necessary to improve the neural network predictions in order to achieve greater correlation between the measurement groups.

This abstract was presented at the 2019 ARVO Imaging in the Eye Conference, held in Vancouver, Canada, April 26-27, 2019.

 

Fig. 1: Results of algorithm. (a) Corner analysis with borders in green; (b) probabilities from the deep neural network; (c) results of algorithm on binarized network outputs with H=hex.

 

Fig. 2: Bland Altman plot of HEX measures between the CAS/EB software manual analysis and our shape detection algorithm using the automatic segmentations.