The purpose of our study is to apply a neural-network technique to detect changes of trabecular patterns due to osteoporosis on skeletal radiographs. Regions of interest (ROIs) (64x64pixels,6.4x6.4mm2) near the center of the vertebral body were extracted from digitized bone radiographs. After the background-trend correction was performed by a curve fitting technique, texture measures were calculated, such as standard deviations (total, horizontal, and vertical directions), maximum and minimum pixel values, and conditional entropy, which were used as input data to the input layer of the neural network. The network was trained with a back-propagation algorithm to discriminate textural differences between the normal and the diseased cases (2 outputs), or between the normal, early stage,1st stage,2nd stage and 3rd stage (5 outputs). A data base of 44 bone radiographs was collected. Half of the cases was used as the training set, and the other half as the testing set. The neural networks with 2 outputs and 5 outputs correctly classified about 86% and 66% of the cases at maximum, respectively. It was found to be most difficult to detect the early-stage image data; the correct recognitions of about 95% and 76% at maximum were obtained for 2 outputs and 5 outputs cases, respectively. Our results suggest that the neural-network analysis is useful to discriminate textural variations of trabecular patterns and is effective for computer-aided diagnosis system of osteoporosis.
View full abstract