Purchase this article with an account.
Juan P. Vigueras-Guillén, Jeroen van Rooij, Hans G Lemij, Lucas J. van Vliet, Koenraad Arndt Vermeer; Accuracy of a deep learning approach for corneal endothelium biomarker estimation in ultrathin-DSAEK images. Invest. Ophthalmol. Vis. Sci. 2019;60(9):3814.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
After keratoplasty, the in-vivo image acquisition of the corneal endothelium is problematic due to the rough endothelial surface and the graft-recipient interface, which affects the optic path of the microscope and results in images degraded by illumination artifacts (Figure 1, left). This makes the segmentation of the endothelial cells – necessary to assess the biomarkers: cell density (ECD), cell size variation (CV), and hexagonality (HEX) – a complicated task. We evaluated an automated framework to estimate such biomarkers.
301 central cornea images from 41 patients who underwent ultrathin-DSAEK surgery were captured at 1, 3, 6, and 12 months post-surgery by using a Topcon SP-1P specular microscope. We designed a completely automated cell segmentation process by employing two deep learning approaches: one to find cell edges (Vigueras-Guillén et al., BMC Biomedical Engineering, 2019) and another to detect the area where cells are reliably detected (Vigueras-Guillén et al., SPIE Medical Imaging, 2019). Based on the resulting segmentation (Figure 1, right), the biomarkers were calculated and compared against the microscope’s estimates and a ground truth (annotated by an expert).
Topcon failed to produce a cell segmentation in 30% of the images, and HEX could not be estimated in 71% of the images due to insufficient segmented cells. In contrast, our approach failed to segment 3% and estimate HEX in 11% of the images (the ground truth was unable to estimate HEX in 7% of them). We computed the average relative error, assigning an error of 100% for the non-estimated biomarkers, achieving a statistically significant difference in favor of our approach in all cases (Figure 2; p<0.0001, paired Wilcoxon test). The average relative error (only considering the estimated biomarkers) was 2.6% in ECD, 6.1% in CV, and 8.0% in HEX for our method, and 9.3% in ECD, 20.0% in CV, and 16.6% in HEX for Topcon.
The estimation error decreased over time, suggesting less illumination artifacts. Our approach is more robust against such artifacts. Even for the best quality images where Topcon could detect cells, our method yields a 2-3 times smaller error.
This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.
Six images (left) and the deep learning segmentation (right). Good quality images (A) are uncommon. Problems to focus the endothelium (C-G) is the norm. Rough corneas create areas out of focus (I-K).
Relative error in biomarkers.
This PDF is available to Subscribers Only