Purpose
To develop a fully automatic, reliable cell detection system for corneal endothelium images from clinical specular microscopy. The system should eventually allow the measurement of the morphometric parameters that provide an objective clinical assessment of corneal endothelium (density, polymegethism, polymorphism). To this purpose, a new cell segmentation method based on a genetic algorithm was developed.
Methods
Thirty images of corneal endothelium were acquired with a specular endothelial microscope (SP-3000P, Topcon Co., Japan) from both healthy and pathological subjects. The region of interest covered an area of about 0.1 mm2, including on average 220 cells. Each cell contour is detected using a genetic algorithm, a method for solving optimization problems based on a selection process that mimics biological evolution. It randomly modifies individuals from the current population to produce the children for the next generation and over successive generations the population evolves toward an optimal solution. For the specific application, a small set of vertexes (i.e. individuals) forming regular hexagons is used as starting population. At each step, the location of each vertex is randomly modified. Thus, the initial regular hexagons evolve into polygons with possibly different number and positions of vertexes, with each vertex reliability evaluated by considering both its correspondence with the actual image (pixels intensity) and the regularity of the polygons. The entire population of vertexes forms a set of polygons that fit the underlying cells contours. The number of polygons is step-wise increased so as to iteratively detect the contour of all the cells in the image (Figure 1). From these contours the morphometric parameters of clinical interest can then be easily computed.
Results
The preliminary results obtained for the contour detection (see example in Figure 1) show that the segmentation of the corneal endothelial cells provided by the proposed method is in good agreement with the ground truth segmentation, obtained with a careful, time-consuming manual analysis.
Conclusions
The proposed totally automatic algorithm appears capable of reliably obtaining the cell contour in regions containing also hundreds of cells. This will allow the straightforward computation of all the important morphometric parameters used in clinical practice.