Components of image intensive diagnosis. Codifying visual features for effusion cytology

Components of visual assessment in the diagnosis of effusions were analyzed using relative operating characteristic. Diagnostic performance in the assessment of malignancy and the specification of metastatic origin was measured for two expert cytologists. The component of performance attributable to feature interpretation was measured in protocols which minimized the effects of clinical information and visual search in the decision process. Feature interpretation, as a process, contributed significantly to the evaluation of malignancy and marginally to the specification of metastatic origin. For each of these diagnostic tasks, the process of feature interpretation was codified in the construction of explicit models. The expert cytologists were asked to define a set of localized visual features that incorporate essential visual elements for diagnosis. These features were evaluated for a set of test cases, and regression models were constructed defining malignancy and metastatic origin. Relative operating characteristic analysis indicated that the predictive value of the models for diagnosis was very similar to the component of human performance attributable to feature interpretation.