PROBABILISTIC NEURAL NETWORK CLASSIFIER VERSUS MULTILAYER PERCEPTRON CLASSIFIER IN DISCRIMINATING BRAIN SPECT IMAGES OF PATIENTS WITH DIABETES FROM NORMAL CONTROLS

摘要

The aim of this study was to compare the performance of the probabilistic neural network (PNN) classifier with the multilayer perceptron (MLP) classifier, in an attempt to discriminate between patients with diabetes mellitus type Ⅱ (DMII) and normal subjects using medical images from brain single photon emission computed tomography (SPECT). Features from the gray-level histogram and the spatial-dependence matrix were generated from image-samples collected from brain SPECT images of diabetic patients and healthy volunteers, and they were used as input to the PNN and the MLP classifiers. Highest accuracies were 99.5% for the MLP and 99% for the PNN and they were achieved in the left inferior parietal lobule, employing the mean value and correlation features. Our findings show that the MLP classifier outperformed slightly the PNN classifier in almost all cerebral regions, but the lower computational time of the PNN makes him a very useful classification tool. The high precision of both classifiers indicate significant differences in radio-pharmaceutical (99mTc-ECD) uptake of diabetic patients compared to the normal controls, which may be due to cerebral blood flow disruption in patients with DMII.