A study of the automated prescreening algorism for cancer cytology of the uterine cervix

This study was undertaken to develop an automated prescreening algorism for cervical cytology of squamous cell carcinoma of the uterus. Routine, vaginal smears stained with the Papanicolaou method were used. The squamous cells were classified into normal and abnormal ones. The former included normal squamous cells and those with a slight degree of atypia which appeared in inflammation and mild dysplasia. The latter included those from lesions worse than a moderate degree of dysplasia. By using a linear discriminant function with nuclear area (NA) and N/C ratio as parameters, the misclassification rate of abnormal cells resulted in 4.6%, when that of normal ones was 50%.By adding nuclear mean density (NMD) to these parameters, the former resulted in 3.3%, when the latter was 50%.Next, squamous cells were classified into normal, abnormal and unclassified ones with NA and N/C ratio and the unclassified cells were reclassified into normal and abnormal ones with NA and NMD (Hierarchic classification algorism). The misclassification rate of abnormal cells resulted in 3.3%, when that of normal ones was 28%.Applying “Nuclear Area with High Density (NAH) ”, instead of NMD, to the hierar chic classification algorism, the former resulted in 2.3%, when the latter was 15% (Cell type independent hierarchic classification algorism). The NAH is a new parameter defined as a nuclear area denser than mean value for minimal and maximal density and is considered to be related to the chromatin pattern. For further improving the result, the unclassified cells were classified in basal, parabasal and intermediate types with NA and N/C ratio and, in each type, they were reclassified into normal and abnormal cells with NA and NAH (Cell type dependent hierarchic classification algorism). The misclassification rate of abnormal cells resulted in 1.7%, when that of normal ones was 6.7%. Thus, the best result was obtained by the cell type dependent hierarchic classification algorism.