Improving ultrasonographic diagnosis of prostate cancer with neural networks

To improve ultrasonographic diagnosis of prostate cancer, the authors evaluated the performance of an optimized backpropagation artificial neural network (ANN) in predicting an outcome (cancer–not cancer) from recorded information on patients admitted for transrectal ultrasonography (TRUS) performed in our Center. A total of 442 cases with complete information were selected for the study. After preselecting 17 variables (age, PSA, previous clinical diagnosis, and 14 ultrasonographic ones) through univariate analysis, a randomly selected subset of data (50%) was used to train ANNs, and the other subset (50%) was used to test the different models. The ANN achieved up to 81.82% of positive predictive value and up to 96.95% of negative predictive value vs. 67.18% and 90.97%, respectively, when compared with those obtained with logistic regression. Results and possible future practical applications are further discussed.