Automatic segmentation with guided correction of post-keratoplasty corneal endothelial cell images and predictive feature extraction.

Abstract

Purpose : We developed a deep-learning-based automatic EC segmentation method followed up by a guided correction graphical user interface (GUI) software to generate accurate cell border representation post-keratoplasty. Furthermore, we have extracted new imaging, textural, and cell graph features that potentially predict transplant rejection.

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). We trained a U-Net model to segment ECs prior to a post-processing pipeline producing single pixel-width borders. We developed a guided correction GUI software, allowing users to add or erase cell borders from suspicious cells, which we highlighted based on under- and over-segmentation conditions. After the cell border segmentation are finalized, we extracted common imaging, textural, and cell graph features. These features will help analyze the healthiness of the EC images following various transplants to determine their successfulness.

Results : Our automatic segmentation method was able to delineate the cell borders of 30 held-out test EC images with a Dice coefficient of 0.87 ± 0.17. From the visual analysis study, there were less than 3% of cells within the ground truth region, which required a second review and possible manual editing. Furthermore, over 500 new cells were identified by our automated segmentation, of which only 31% were flagged for manual editing. We designed a guided correction GUI to highlight 167 suspicious cells (under- or over-segmented cells) for editing. From these finalized segmentations we calculated 190 intensity, textural, and cell graph features.

Conclusions : Our automatic segmentation and guided correction GUI software is able to accurately identify the cell borders of EC images post-keratoplasty. Furthermore, the new features we have calculated to evaluate EC health are novel and possibly indicative of future rejection. The graph features analyze the arrangement of cells in the endothelium which may be indicative of future rejection, since as the endothelium deteriorates, cells start to morph, merge, and create disarray.

This is a 2020 Imaging in the Eye Conference abstract.

 

Examples of post_DSAEK EC image undergone manual and automatic segmentation.

 

Screenshot of semi-automated segmentation GUI software.