Cone beam filtering using artificial neural networks

The authors introduce a possible method to determine and implement cone beam filters that perform in-slice and axial filtering through the use of an artificial neural network. In-plane and interslice filtering are accomplished separately in order to decrease the complexity of this initial neural network experiment. This particular procedure utilized supervised training with the modified delta rule that minimized the mean-squared error through a gradient descent. The in-slice filtering problem was to test if the network could learn the ramp filter for a cone beam geometry. This study used simulated Monte Carlo data that represented a geometry of a point source located off the axis of rotation. Though preliminary data were promising, it was not possible to determine a general axial filter. This is due to insufficient sampling in the axial direction by the cone beam geometry.<>