Differentiation of intracardiac tumors and thrombi by echocardiographic tissue characterization: comparison of an artificial neural network and human observers.

The feasibility of classifying ultrasound images of intracardiac tumors and thrombi with a neural network-based algorithm was compared with the performance of experienced echocardiographers. The neural network used statistical descriptors of the apparent echocardiographic texture of the masses, and the blinded echocardiographers were given photographic prints of enlarged regions of interest without clinical data. The network classified 66% of the images correctly and the echocardiographers, 83%. The network and echocardiographers agreed in 88% of the images. Human observers usually base their classification of intracardiac masses on clinical data. The echocardiographic texture of tumors is quantitatively different from that of thrombi. This difference can be recognized by a neural network and potentially be useful in assisting with the diagnosis when clinical clues are insufficient.