TU‐D‐304A‐02: Dose Reduction in CT Using Non‐Convex Prior Image Constrained Compressed Sensing NC‐PICCS

Purpose: To demonstrate the dose reduction potential in CT using a novel algorithm (NC‐PICCS) based on a non‐convex l₀ homotopic approximation. Methods: We generalized the Prior Image Constrained Compressed Sensing (PICCS) algorithm to solve an l₀‐quasinorm by homotopic approximation using continuation on the parameter p (l_p‐norm) which starts at p=1 and ends at p ≅ 0. The algorithm was validated using both computer simulations and in‐vivo animal studies. A pig perfusion study dataset was used to show the advantages of this novel method over existing compressed sensing approaches. Results: the NC‐PICCS method allows exact image reconstruction to be achieved from fewer projections than with methods such as standard CS or PICCS, both of which employ a convex l₁ norm. For the synthetic data used, we were able to reconstruct the image with as few as 4 projections when a prior image was available in contrast to 12 projections without a prior image. When the method was applied to the in‐vivo pig perfusion data the number of projections required for exact reconstruction was about 20. Conclusions: the NC‐PICCS method provides a framework to highly reduce the dose in CT time series studies. Although there is no theoretical guarantee of finding a global minimum due to the non‐convex nature of NC‐PICCS, substantial empirical evidence suggests that it performs better than previously reported compressed sensing methods for CT reconstruction in practice.