A vision machine is able to measure a nearly unlimited number of parameters when analyzing an image, but we completely agree with Keenan  who pointed out that the development of a mechanical vision system is complicated.
The purpose of this work was to limit the number of parameters by retaining only those values which are as close as possible to the parameters taken into account when a pathologist analyses and grades a cervical biopsy. With a specific algorithm we were able to record automatically 6 reliable parameters and to establish a cervical score for grading objectively a cervical biopsy. We took into account the original classification proposed by Richart  which is still the gold standard in the field and pointed out the epithelial differentiation parameter.
Image captures were done with a ×25 objective which suits perfectly to our optical system, allowing maximum vision of total epithelium height, especially in the normal group. The application of contrast staining appeared to be a valuable procedure in the segmentation process. Segmentation is a major problem in image analysis, as the nuclei are often superimposed, particularly in high grade lesions and cancer. Up to now there is no valid solution, because the manually segmenting of clusters is time consuming and subject to errors, particularly in high grade lesions and cancer. After several attempts to achieve a fully automatic process, it appeared that a manual or semi-automatic procedure was necessary to reach better segmentation before the automatic measurement of the pertinent parameters. This is a general problem in image analysis and represents an important issue for the standardization of image quality, as pointed out by Kayser [14, 15].
Voronoï tessellation is based on nuclei segmentation and determines on each nucleus the zone of influence, which is a mathematical morphology item that suits well with biological observation, because the nucleus exerts an "influence" on the surrounding cytoplasm. Therefore, we can relate the zone of influence area to cytoplasm area. As proposed by Serra (Personal Communication), the Delaunay triangulation was then drawn from the geodesic center of each zone of influence, this procedure allowing a more accurate triangulation because it takes into account the nearest neighbors and eliminates false relation of 2 cells being far away from each other, when there is convexity in the epithelium for example.
The Delaunay triangulation measures the mean triangle edge length which was used to define the dilation/erosion coefficient. By combining this value with the mean value of nuclei area it became possible to automatically measure differentiated epithelium. An alternative for the measurement the dilation/erosion coefficient consist in the application of the minimum spanning tree technique giving rise to the minimum distance between the nuclei's nearest neighbors . A specific algorithm has to be developed in our vision machine in order to check which is the adequate technique. Our procedure, using automatic dilatation/erosion makes it easier in cases of convexity or irregularities of the basal lamina or papillomatosis, which represents most of the cases encountered in routine pathology. Considering percentage differentiation, correlation with Richart's classification was found, though normal epithelium is made up of limited number of basal cells, not considered as differentiated upon microscopic observation, but displaying a small degree of differentiation with the vision machine.
Cell maturation in the upper part of the differentiated epithelium represents an important factor in the grading, especially in cases of koïlocytosis. Our staining procedure is based on iron-haematoxylin treatment. By increasing the nucleus contrast with our modified technique, although haematoxylin is not stochiometric per se, the intensity of the nucleic staining, which is taken into account in routine microscopy, can bring up an interesting data. When establishing the histogram of all nuclei on one end and the upper nuclei on the other end, the histogram ratio gives an adequate appreciation of nucleic chromaticism, independent of specimen thickness and staining intensity.
Nucleo-cytoplasmic ratio is an important data from the spatial point of view. In normal epithelium near the surface, the cytoplasm is very large and the nucleus becomes smaller during normal maturation. As the histological section is 2.5 μm thick, only a few nuclei are present with their surrounding cytoplasm, leading to a false estimation of nucleo-cytoplasmic ratios. Data are nevertheless comparable because they are objectively done by the machine and are reproducible.
We did not consider mitoses which is an important parameter in the classification. This can be achieved when images are captured with a ×1000 magnification . Other parameters such as DNA ploïdy can be measured by image analysis [17, 18], but they represent a more complicated approach and are time consuming. We also did not take into account the immuno-histochemistry with Ki65 and P16, which can represent an additional parameter . Comparing data from the 4 pathologists, a good correlation was observed with the machine in 79% of the cases, with consensus cases taken into consideration for the selection of the 4 groups submitted to the vision machine. The data produced by automatic calculation display some statistical significance among the 4 groups, but the low number of cases in each group may be a limitation. Twenty six more images were added to the 86 previously analyzed, and were submitted to the 4 pathologists in a confrontation meeting without regard to the previous results. There was 77% correlation between the consensus diagnosis and the vision machine results which confirms the results of Keenan .
The use of mathematical morphology makes our algorithm transposable to any system based on that image analysis system, after proper calibration of the optical system specific to each user. The staining procedure, the microscopic field illumination and the automatic measures with our system were standard and gave reproducible results, these parameters being considered as very important .
The cervical score we proposed presents a quasi-linear progression from normal to high grade lesions and fits closely to Richart's classification. The problem however is to determine an adequate cut-off value for the decision to treat or not to treat a patient.
Though it may not always be comfortable to accept that a machine can be able to grade a CIN, we hope that our contribution to the proposal for a reliable scoring will help pathologists to deliver objective diagnoses, further more in experimental studies where objectivity and quantification are requested. So far, we consider that the vision machine is not able to produce a diagnosis per se, the selection of the optical field to be analyzed remaining operator dependent. Nevertheless it can help the pathologist to accurately establish the most valuable diagnosis. Although some artefacts may occur, precision seems fairly good, which is hardly the case when grading is done by routine observation on microscopic slides, as inter and intra variability is a limiting factor. The present investigation should be considered as a contribution to total morphologic analysis of any cervical lesion by measuring objective parameters, with special reference to epithelial differentiation.