Investigation of and preliminary results for the solution of the inter-observer variability problem using fine needle aspirate (FNA) data

Abstract

The paper provides a preliminary evaluation of the accuracy of computer aided diagnostics (CAD) in addressing the inconsistencies of inter-observer variance scoring. The inter-observer variability problem, in this case, relates to different cytopathologists and radiologists at separate locations scoring the same type of samples differently using the same methodologies and environmental discriminates. Two distinctly different FNA data sets were used. The first was the data collected at the University of Wisconsin (Wolberg data set) while the other was a completely independent one defined and processed at the Breast Cancer Center, University Health Center at Syracuse (Syracuse data set). Two computer aided diagnostic (CAD) paradigms were used: the evolutionary programming (EP)/probabilistic neural network (PNN) hybrid and a mean of predictors model. Four experiments mere performed to evaluate the hybrid. The fourth experiment, k-fold crossover validation, resulted in a 91.25% average classification accuracy with a .9783 average Az index. The mean of predictors model was used to verify the results of the more complex hybrid using both the fraction of missed malignancies (Type II errors) and fraction of false malignancies (Type I errors). The EP/PNN hybrid experiments resulted in a 3.05% mean value of missed malignancies (Type II) and a 5.69% mean value of false malignancies (Type I errors) using the k-fold crossover studies. The mean of predictors model provided a.429% mean Type II error and a 4.09% mean Type I error.