Compressed Sensing Based Four Dimensional Cone-Beam Computed Tomography Reconstruction by Continuation Accelerated Nesterov's Descent
H Zhang*, J Sonke, Netherlands Caner InstituteAmsterdamTU-A-213CD-7 Tuesday 8:00:00 AM - 9:55:00 AM Room: 213CD
Compressed sensing based cone-beam reconstruction is introduced to reduce streaking artifacts due to under-sampling in four dimensional (4D) cone-beam computed tomography (CBCT). This reconstruction, however, is typically time-consuming, especially for 4D CBCT. We propose a novel approach, which introduces a first-order methods, called continuation accelerated Nesterov's descent, to improve convergence rate.
We extract the respiratory signal directly from projections, sort them into ten subsets, and reconstruct each subset into a 3D CBCT by compressed sensing based total-variation minimization. The reconstruction iteration is solved by Nesterov's descent algorithm. In each loop, we adapt the step size by sufficient reducing global Lipschitz constant, and update variants by gradient-mapping parameters to proximal points. The proposed method is assessed by a thorax phantom and a patient, which are scanned with a CBCT scanner. The reconstructed data is compared among FDK algorithm, projection onto convex sets (POCS) which is a steepest decent based optimization, and our method.
A region of homogeneous voxel values is manually selected, and the root-mean-square-error between CBCT and CT in this region is defined as streaking artifacts. It's reduced 27.2% in phantom data, when our method is compared to traditional FDK methods for 2 minute scan. Correlation ratio between CBCT and CT is increased for 10.2% when our method is compared to FDK for patient data. For the convergence rate, our method only needs 25 iterations to get the same total-variation residual of POCS at 250 iterations. Also for other cut-offs, it seems an order of magnitude faster.
Our proposed method decreases the streaking artifacts and convergence time in 4D CBCT reconstruction. The image quality assessment suggests that our method has improved image quality and better correlation.