Diagnosing mitral and tricuspid stenosis with the help of artificial neural networks built on the fast Fourier transformation sonogram

Mitral and tricuspid stenoses are currently diagnosed through the use of Doppler Sonograms. However, the final accurate diagnosis depends on the physicians' interpretation and experience and this sometimes results in false diagnosis. In this study, we have facilitated artificial neural networks (ANN) that will not only simplify the diagnosis but also enable the physician to make a quicker judgment about the presence of stenosis, in confidence. The ratios of the three points (First Systolic Peak, Endpoint of Diastole, Second Systolic Peak) in the M like curve of the sonogram are used as inputs to our ANN. We have chosen Levenberg-Marquart training algorithms to train the multi layer perceptron (MLP) structure of ANN, since the least mean square error of 8/spl times/10/sup -5/ and shortest convergence time of 1 sec were achieved through this method after comparing eight different algorithms. Our system is tested on 14 patients. Four of the six tricuspid patients' and six of the eight mitral patients' data are allocated for training purposes and the remaining two tricuspid and two mitral patients were tested. The testing results were found to be compliant with physicians' findings regarding the presence of stenosis.