Prediction of Locations in Medical Images Using Orthogonal Neural Networks

An orthogonal neural network (ONN), a new deep-learning structure for medical image localization, is developed and presented in this paper. This method is simple, efficient, and completely different from a convolution neural network (CNN). The diagnostic performance of ONN for detecting the location of pneumothorax in chest X-rays was assessed and compared to that of CNN. An area under the receiver operating characteristic (ROC) curve (AUC) of 0.870, an accuracy of 85.3%, a sensitivity of 75.0%, and a specificity of 86.5% were achieved; the ONN outperformed the CNN. The diagnostic performance of the ONN with a sigmoid activation function for all the nodes obviously outperformed the ONN with the rectified linear unit (RELU) activation function for all the nodes other than the output nodes. In addition, by applying ONN and CNN to predict the location of the glottis in laryngeal images, we achieved accurate and adjacent prediction rates of 70.5% and 20.5%, respectively, with the ONN. The prediction accuracy of the ONN was compared favorably with that of the CNN. Compared to a CNN, an ONN required only approximately 10% of the computations using a CNN trained on images with an input resolution of 256 × 256 pixels. A fully-connected small artificial neural network (ANN), selected by comparing the test results of several dozens of small ANN models, achieved the best location prediction performance on medical images. This study demonstrated that an ONN can be used as a quick selection criterion to compare ANN models for image localization since an ONN performed well compared decently with the selected ANN model.