An efficient framework for high-quality low-dose CT reconstruction and reference-based image restoration.

摘要

CT imaging procedures have been shown to considerably increase the medical radiation dose to patients, giving rise to cancer. As a result, low dose imaging modalities have gained much attention recently. However, this renders traditional CT reconstruction processes no longer sufficient. Iterative reconstruction algorithms using numerical optimization paradigms are better suited, but they suffer from (1) expensive computation, (2) problems with the selection of optimal parameters to simultaneously optimize speed and (3) poor reconstruction quality.;To cope with these problems, we have made several contributions to low-dose CT. First, we have devised a GPU-accelerated ordered subset iterative CT reconstruction algorithm (OS-SIRT) with regularization and effective parameter learning. We generalized two algebraic algorithms (SIRT, SART) to an ordered subset scheme which balanced the speed of computation and the rate of convergence of these algorithms. Second, we mapped the computation to GPUs, achieving remarkable performance gains. Third, for high-quality reconstruction, we introduced four filters for denoising and streak artifact reduction, i.e., bilateral, trilateral, non-local means (NLM) and optimal adaptive NLM, all of which are popular in computer vision. We have used these filters within an interleaved CT reconstruction regularization pipeline and found that they compare favorably with the traditionally used TVM algorithm. Fourth, to overcome the difficulties with optimal parameter tuning within our algorithm and for any parameter-dependent applications, we devised two parameter-learning approaches---exhaustive benchmark testing and multi-objective optimization---and allow user interaction via an interactive parameter space visualization tool. We then generalized our framework to Electron Tomography. Our fifth contribution is a scheme that broadens the low-dose image restoration capability of traditional NLM filtering to also include high-dose reference images. We developed two variants. The first variant uses a prior scan of the same patient when available. The second variant generalizes this concept to a database of images of other patients to learn the reference images. Our experiments show that this scheme has vast potential to restore the quality of low-dose CT imagery.