DenseUNets with feedback non-local attention for the segmentation of specular microscopy images of the corneal endothelium with guttae

Abstract

To estimate the corneal endothelial parameters from specular microscopy images depicting cornea guttata (Fuchs dystrophy), we propose a new deep learning methodology that includes a novel attention mechanism named feedback non-local attention (fNLA). Our approach first infers the cell edges, then selects the cells that are well detected, and finally applies a postprocessing method to correct mistakes and provide the binary segmentation from which the corneal parameters are estimated ( cell density [ECD], coefficient of variation [CV], and hexagonality [HEX]). In this study, we analyzed 1203 images acquired with a Topcon SP-1P microscope, 500 of which contained guttae. Manual segmentation was performed in all images. We compared the results of different networks (UNet, ResUNeXt, DenseUNets, UNet++) and found that DenseUNets with fNLA provided the best performance, with a mean absolute error of 23.16 [cells/mm 2 ] in ECD, 1.28 [%] in CV, and 3.13 [%] in HEX, which was 3-6 times smaller than the error obtained by Topcon's built-in software. Our approach handled the cells affected by guttae remarkably well, detecting cell edges occluded by small guttae while discarding areas covered by large guttae. Overall, the proposed method obtained accurate estimations in extremely challenging specular images.

Deep learning segmentation of endothelial cell images using an active learning paradigm with guided label corrections

“…In the application of corneal EC image segmentation, deep learning studies only cover methods to automatically analyze EC images using U-Net and its variations. 11,[13][14][15][16] A few studies in other applications, such as liver and spleen segmentation in CT images, have reported human-in-the-loop and reinforcement learning approaches as means to create large quantities of ground truth segmentations. 26 Our study builds on the literature by exploring a different deep learning network, DeepLabV3+, and its enhanced segmentation performance on post-DMEK EC images compared to U-Net.…”

“…Limitations arose when segmenting larger cells and images with nonuniform intensities or low contrast. Most recently, Vigueras-Guillén et al [13][14][15][16] developed enhanced U-Net models with combined region of interest and edge detection, dense blocks, feature map rotations and reflections, and feedback nonlocal attention blocks.…”