Automated corneal endothelium image segmentation in the presence of cornea guttata via convolutional neural networks

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

“…In recent years, several new approaches to estimate the corneal endothelial parameters have been proposed. Up to 2018, the methods were based on image processing techniques and classic machine learning [13][14][15][16][17] , but from 2018 onward, a significantly large number of new approaches based on deep learning have been presented [18][19][20][21][22][23][24][25][26][27][28][29][30][31][32] . Overall, these methods have shown a good performance.…”

“…However, it is worth noting a few details: (i) some methods (mainly from the pre-deep learning era) evaluated the estimation of the three corneal parameters but the images were from healthy eyes and there was either a manual selection of the region of interest (ROI) or the cells were visible in the whole image [14][15][16][17][18]21,23 ; (ii) some of the first deep learning methods simply evaluated the capacity of neural networks to perform an accurate segmentation, either in healthy 19,22 or unhealthy corneas 26 , without estimating any corneal parameter, which avoids the non-trivial problem of refining the raw segmentation for the purpose of obtaining an accurate biomarker estimation; (iii) other publications only focused in estimating ECD (or the number of cells) in healthy cases 29,30 and images with guttae 20,27,31 , some including a method to select the ROI (although this part is often unclear); and (iv) our previous work is, to the best of our knowledge, the only fully-automatic method to estimate all three parameters in all types of images (heavily blurred 24,25 and also with some guttae 28 ). Among the publications dealing with guttae, a quick visual inspection is enough to perceive the inaccurate segmentation around the guttae, where partial cells occluded by the guttae are included as full cells 27,31 ; in contrast, our previous work 28 has shown better results but still failed in cases of very advanced disease. Therefore, the accurate segmentation of images in the presence of guttae is still an unsolved problem.…”

Computerized Detection of Limbal Stem Cell Deficiency from Digital Cornea Images

“…The following features: Horizontal and vertical Visible Iris Diameter, curvature radius, spherical aberration and solidity are measured using our segmentation method and compared to other methods in the literature. Solidity is an important unique feature of corneal surface [14]. These features characterize corneal surface structure where the Horizontal Visible Iris Diameter is longer than the vertical Visible Iris Diameter for a healthy cornea.…”

“…They measure their technique performance in CEC images with different corneal diseases. Sierra [14] performed cell segmentation from specular images of both healthy and LSCD corneas. Cell segmentation is processed through supervised segmentation of digital images.…”

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

“…However, it is worth noting a few details: (i) some methods (mainly from the pre-deep learning era) evaluated the estimation of the three corneal parameters but the images were from healthy eyes and there was either a manual selection of the region of interest (ROI) or the cells were visible in the whole image (14-18, 21, 23); (ii) some of the first deep learning methods simply evaluated the capacity of neural networks to perform an accurate segmentation, either in healthy (19,22) or unhealthy corneas (26), without estimating any corneal parameter, which avoids the nontrivial problem of refining the raw segmentation for the purpose of obtaining an accurate biomarker estimation; (iii) other publications only focused in estimating ECD (or the number of cells) in healthy cases (29,30) and images with guttae (20,27,31), some including a method to select the ROI (although this part is often unclear); and (iv) our previous work is, to the best of our knowledge, the only fully-automatic method to estimate all three parameters in all types of images (heavily blurred (24,25) and also with some guttae (28)). Among the publications dealing with guttae, a quick visual inspection is enough to perceive the inaccurate segmentation around the guttae, where partial cells occluded by the guttae are included as full cells (27,31); in contrast, our previous work (28) has shown better results but still failed in cases of very advanced disease. Therefore, the accurate segmentation of images in the presence of guttae is still an unsolved problem.…”