Fully convolutional architecture vs sliding-window CNN for corneal endothelium cell segmentation

Abstract

Background: Corneal endothelium (CE) images provide valuable clinical information regarding the health state of the cornea. Computation of the clinical morphometric parameters requires the segmentation of endothelial cell images. Current techniques to image the endothelium in vivo deliver low quality images, which makes automatic segmentation a complicated task. Here, we present two convolutional neural networks (CNN) to segment CE images: a global fully convolutional approach based on U-net, and a local sliding-window network (SW-net). We propose to use probabilistic labels instead of binary, we evaluate a preprocessing method to enhance the contrast of images, and we introduce a postprocessing method based on Fourier analysis and watershed to convert the CNN output images into the final cell segmentation. Both methods are applied to 50 images acquired with an SP-1P Topcon specular microscope. Estimates are compared against a manual delineation made by a trained observer. Results: U-net (AUC = 0.9938) yields slightly sharper, clearer images than SW-net (AUC = 0.9921). After postprocessing, U-net obtains a DICE = 0.981 and a MHD = 0.22 (modified Hausdorff distance), whereas SW-net yields a DICE = 0.978 and a MHD = 0.30. U-net generates a wrong cell segmentation in only 0.48% of the cells, versus 0.92% for the SW-net. U-net achieves statistically significant better precision and accuracy than both, Topcon and SW-net, for the estimates of three clinical parameters: cell density (ECD), polymegethism (CV), and pleomorphism (HEX). The mean relative error in U-net for the parameters is 0.4% in ECD, 2.8% in CV, and 1.3% in HEX. The computation time to segment an image and estimate the parameters is barely a few seconds. Conclusions: Both methods presented here provide a statistically significant improvement over the state of the art. U-net has reached the smallest error rate. We suggest a segmentation refinement based on our previous work to further improve the performance.

Convolutional neural network-based regression for biomarker estimation in corneal endothelium microscopy images

“…Previously, we presented a fully automated framework to estimate the endothelial biomarkers from specular microscopy images (Figure 1). This framework is subdivided into four different methods: a CNN based on a Dense U-net to segment endothelial cells using an intensity image as input and providing an edge probability image as output (CNN-Edge) [9], [10]; a CNN based on a Dense U-net to segment the region of interest where cells are correctly detected using the edge probability image as input (CNN-ROI) [10]; a postprocessing method based on Fourier analysis and watershed that combines both output images and yields the binary edge image [9], [11]; and a machine learning approach based on Support Vector Machines (SVM) that removes potential false edges from the binary edge image [8]. The biomarkers could then be estimated from the final binary image.…”

“…These image-based biomarkers can be easily estimated if the cell boundaries are identified in the image, therefore image segmentation has been the approach most commonly employed for solving this task [6], [7], [8]. In previous work, we achieved state-of-theart results by employing a CNN U-net to segment the cell boundaries [9], but in order to make the process completely automated we developed a dense U-net to infer the area in the image (region of interest, ROI) in which the biomarkers could be reliably estimated [10]. Furthermore, it is not a trivial task to combine both CNN output images (the edge probability image and the ROI image) and transform them into the final binary edge image to perform the biomarker estimation.…”

“…Indeed, any small error or discontinuity in an edge could cause a significant error in the resulting estimates. This step was called post-processing in our previous work [9].…”

DBT Masses Automatic Segmentation Using U-Net Neural Networks

“…Elboushaki et al validated that the CNN model can recognize fine mammographic features [23]. Vigueras-Guillen et al first proposed a full-convolution network for semantic segmentation, replacing the conventional full-connection layer in the CNN with the convolution layer to obtain a rough label graph and then using the deconvolution layer to sample the rough label graph to achieve the classification results of each pixel [24]. Ciresan et al used patches of 101 × 101 pixels to train a CNN for mitosis detection in breast cancer histology images, who won the ICPR 2012 Mitosis Detection Contest with F1-score of 0.782 [25].…”

Semantic Segmentation of Endangered Tree Species in Brazilian Savanna Using Deeplabv3+ Variants

“…In this case, the image is split into several snips or patches of the same size and then fed into a traditional CNN to classify the central pixel of each patch in a certain category (Farabet et al, 2012). One important drawback of this approach is its high computational complexity (Vigueras-Guillén et al, 2019). The second approach comprises the efficient Fully Convolutional Networks (FCNs), in which the image classification is performed at a pixel level using an end-to-end network.…”