We are developing a computerized system that can aid in the physician's diagnosis in the detection and classification of coronary artery diseases. The technique employs a multi-layer feedforward neural network with a backpropagation algorithm to analyze Tl myocardial SPECT bull'seye images. In this study, the effects of image preprocessing, matrix sizes, and some neuro parameters have been investigated by use of EXTENT images. Tests with images of 32x32 matrix along with inverse processing resulted in best recognition rate, but 16x16 matrix can be considered enough due to several reasons, such as reduced training time and data volume and improved result with an inclusion of SEVERITY images. The number of hidden units was changed and it was found that the onehundred units for our case are reasonable. Increasing the hidden layer from a single to double with 10x10 or 20x20 cells was effective in terms of training time at least with the same correct recognition rate. Our study indicated that the order of image data for the training process has to be considered for improvement of the recognition.

View full abstract