Neural network scatter correction technique for digital radiography

A scatter correction technique based on artificial neural networks is presented. The technique utilizes the acquisition of a conventional digital radiographic image, couple with the acquisition of a multiple pencil beam ('micro-aperture') digital image. Image subtraction results in a sparsed sampled estimate of the scatter component in the image. The neural network is trained to develop a causal relationship between image data on the low-pass filtered open field image and the sparsely sampled scatter image, and then the trained network is used to correct the entire image (pixel by pixel) in a manner which is operationally similar to but potentially more powerful than convolution. The technique is described and is illustrated using clinical 'primary' component images combined with scatter component images that are realistically simulated using the results from previously reported Monte Carlo investigations. The results indicate that an accurate scatter correction can be realized using this technique.