Digital image classification and detection using a 2D-NARX model

Digital image processing is shifting the information analysis paradigms in a variety of systems whereas that it is becoming highly viable. Medicine has made the most of image processing by enhancing decision-making through computer-aided diagnosis (CAD) systems. CAD image models support medical inferences by extracting key visual features and classifying regions of interest. Linear parametric system identification models, typically used for time series forecasting such as the auto regressive moving average (ARMA) model have been adopted as CAD image procedures with encouraging results. But this perspective can be enhanced by incorporating flexible-order system identification models capable to detect corrupting features that might be shortly described by linear models. In this way, the polynomial nonlinear autoregressive with exogenous inputs (NARX) model is presented along with the forward regression orthogonal least squares (FROLS) algorithm and the k-means clustering method. For the first time the polynomial NARX is presented as a system identification method for digital images and as a CAD procedure for breast cancer detection in images. Experiments with a public database of mammograms show that the method actually detected nonlinearities within images and attained competitive performance measures compared to previous similar methods.