| Abstract: | Digital image analysis was used to extract features from 1,123 abnormal cells in 23 routinely prepared, Papanicolaou-stained cervical smears. All slides examined had a cytologic diagnosis of moderate dysplasia. Seven slides came from patients who eventually progressed to either a severe dysplasia or carcinoma in situ; the other 16 slides came from patients whose dysplasias regressed without evidence of more serious disease. Linear discriminant analysis correctly classified approximately 73% of the cells from the regression group and 66% of the cells from the progression group. Cell features contributing to the majority of variance in the model were the mean optical density of the nucleus, an autocorrelation measure, the mean optical density of the cytoplasm and the nuclear-cytoplasmic ratio. At the patient level, 13 (81.2%) of the 16 slides from the regression group and 6 (85.7%) of the 7 slides from the progression group were correctly classified. These figures yield a sensitivity of 66.7%, a specificity of 92.9%, a predictive value of a progression prognosis of 85.9% and a predictive value of a regression prognosis of 81.2%. The overall efficiency of the model was 82.6%. These preliminary results should encourage further studies for the identification of markers to indicate which patients are at high risk for progression of their cervical dysplasias. |
