Neural networks in ventilation-perfusion imaging. Part II. Effects of interpretive variability.

Abstract

To evaluate the usefulness of a neural network developed by one physician and used by another. Intra- and interobserver variability were analyzed in image categorization of ventilation-perfusion (V-P) scans. This information was used to estimate network performance when it was used by a physician who did not train the network. Network training was optimized by using input parameters that demonstrated both individually high correlations with pulmonary embolism and good reproducibility in multiple interpretations. Potential variability exists in the performance of a network when it is supplied with input data by different physicians. The clinical usefulness of a network depends heavily on the similarity of interpretive styles between the network trainer and the user.