Maximum entropy co-processor for computed tomography

In this paper, we present a neural net co-processor capable of performing computed tomographic image reconstruction. The circuit performs the Radon transformation using a cost function gradient descent method. The unique aspect of this co-processor is the incorporation of informational entropy as a regularizer in the optimization problem. A 10 pixel/spl times/10 pixel array was designed and fabricated in 2 /spl mu/m CMOS technology. Convergence time of the array was less than 5 /spl mu/s. Issues relating to scaling the array to larger sizes are discussed in this paper.<
>