Corneal endothelium image segmentation using feedforward neural network

In this paper the problem of corneal endothelium image segmentation is considered. Particularly, a fully automatic approach for delineating contours of corneal endothelial cells is proposed. The approach produces one pixel width outline of cells. It bases on a simple feedforward neural network trained to recognize pixels which belong to the cell borders. The edge probability (edginess) map output by the network is next analysed row by row and column by column in order to find local peaks of the network response. These peaks are considered as cell border candidates and in the last step of the method via binary morphological processing are linked to create continuous outlines of cells. The results of applying the proposed approach to publicity available data set of corneal endothelium images as well as the assessment of the method against ground truth segmentation are presented and discussed. Obtained results show, that the proposed approach performs very well. The resulting mean absolute error of cell number determination is around 5% while the average DICE measure reaches 0.83 which is a good result, especially when one pixel width objects are compared.