Prostate cancer is the most common malignant disease in American men: In 1991 over 120,000 new cases will be diagnosed and over 30,000 men will die. Though tumor detection depends primarily on physical examination along with biochemical and immunologic tests, imaging tests are required for tumor staging. Magnetic resonance imaging (MRI) is the main candidate for diagnosing invasion of tissues beyond the prostate gland itself and. hence, for determining whether the treatment is to be surgery, radiation therapy, or some merely palliative alternative. A recent study of conventional MRI with a body coil that was made by the Radiology Diagnostic Oncology Group (RDOG) found its accuracy to be lower than desired and expected. The RDOG is now collecting cases imaged by MRI with the newly developed surface coil, which shows greater promise. The objective of the proposed project is to enhance further the accuracy of (surface-coil)MRI so that it can be used in practice with substantial effectiveness. Our approach is to refine and apply techniques that will: (a) identify the perceptual features of MR images that are diagnostically relevant and also the relative importances of those features and (b) aid the image reader in numerically scaling those features, and combining them into an estimate of the probability of the disease stage in question. Previous applications of these techniques to the perceptually similar tasks of diagnosing breast cancer by mammography and diaphanography have shown large increases in diagnostic accuracy. These feature-analytic and decision-aiding techniques will be applied to the set of pathology-proven cases collected in the RDOG study, using the original films. Investigators from the four sites of the RDOG study will participate as advisors on MRI interpretation, on the image data base and imaging tests, and also as expert image readers, working in collaboration with experimental psychologists and radiologists experienced in the enhancement techniques. The series of steps to be followed in a programmatic application of the techniques will include interviews and consensus meetings with the experts, and also perceptual studies and psychometric analyses--in order to define the candidate perceptual features and their numerical scales. The experts then scale the features of proven cases and these data are subjected to statistical analyses in order to establish the minimal but sufficient set of features. In a final step, computer classifiers are developed that accept scale values from image readers of new cases and merge those values into advisory estimates of the probabilities of particular disease stages. Multiple classifiers are developed to treat distinctions between different stages of disease. In the course of this program, several practical and theoretical questions will be examined in the interests of making the feature-analytic and decision-aiding techniques still more powerful. Then an evaluation involving five MRI readers will be conducted to compare their average accuracy in reading the cases in the standard manner with their average accuracy in a reading as enhanced by the perceptual and decision techniques. Accuracy will be measured in terms of the relative operating characteristic, or ROC.
