Cell contour detection in corneal endothelium in-vivo microscopy

In-vivo microscopy of corneal endothelium is a technique routinely adopted in ophthalmic investigation. The main features to be extracted are cell density and morphology. Even if recognizing cell contour and counting cell bodies is common practice for any ophthalmologist, this task has proven to be very difficult in its automation due to the very low image quality. The problem is actually a difficult segmentation problem where the two regions to be identified are cell bodies and cell boundaries. The quality of the image is generally so compromised that conventional and also sophisticated segmentation algorithms found in the literature do not work. The new algorithm we have developed is composed of a first segmentation module, based on a neural network structure with two-dimensional inputs and outputs, whose neuron weights are numerical filters specifically designed for a border extraction problem and derived from the boundary contour system approach (Grossberg et al.). However, gray-level information alone is not sufficient to correctly discriminate between cell bodies and cell boundaries. To cope with the wrong results still present in the segmented image (missing contours or false contours), a second module has been developed, which applies a so called “expert correction”, based on an automatic, multiple-step approach, which includes missing contour recovery and tentative merging or splitting of cell bodies. The preliminary results are very satisfactory, since in a set of more than 300 images of corneal endothelium the difference between the automatic cell count and the human expert one is on average less than 5%