Encrypted login | home

Program Information

Low-Dose Conebeam CT Reconstruction Using Block-Matching 3D-Transform (BM3D) Regularization

Q Lyu

Q Lyu1*, C Yang2 , S Gong2 , Y Xue2 , J Wang2 , D O'Connor1 , T Niu2 , K Sheng1 , (1) UCLA School of Medicine, Los Angeles, CA, (2) Zhejiang University, Hangzhou, Zhejiang,


TH-AB-601-8 (Thursday, August 3, 2017) 7:30 AM - 9:30 AM Room: 601

Purpose: The concerns on the risks of imaging dose to radiotherapy patients from repeated Cone Beam CT (CBCT) motivate the development of low-dose CBCT (LdCBCT). State-of-art Total variation (TV) regularized iterative reconstruction has shown better image artifact reduction and noise control than the analytical filtered backprojection algorithm. However, due to the local gradient operation, TV tends to substantially degrade image resolution with high level of noise presented in LdCBCT. Here we propose a novel iterative reconstruction method using Block-Matching 3D-transform shrinkage (BM3D) regularization, to maximally preserve resolution and suppress noise in LdCBCT.

Methods: To incorporated BM3D transformation into LdCBCT as non-local patch-wise regularization, the iterative reconstruction was formulated as a convex optimization problem with a quadratic data fidelity term regularized by the L1 norm of image spectra in the BM3D transformation domain, where the BM3D transformation matrix is generated based on an initial image via a block-matching step. The convex optimization problem was then solved using the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA). BM3D-regularized LdCBCT (BM3D-LdCBCT) was compared to TV-regularized LdCBCT (TV-LdCBCT) in the reconstruction of down-sampled projection data acquired from a clinical on-board imager.

Results: In the Catphan©504 phantom study using 17% of 610 projections, BM3D-LdCBCT better preserved the line pair contrast compare to TV-LdCBCT at the same noise STD. The modulation transfer function (MTF) of BM3D-LdCBCT is a factor of 2.24 greater than that of TV-LdCBCT at 50% MTF. In the head patient study using 25% of 367 projections, BM3D-LdCBCT increased the spatial resolution by a factor of 2.02 as compared to TV-LdCBCT. The increased image resolution at the same noise level translated into substantially sharper images.

Conclusion: The novel formulation of non-local BM3D as L1-regularization in iterative CT reconstruction is superior to local TV regularization in preserving high spatial resolution while suppressing noise for the clinical LdCBCT.

Funding Support, Disclosures, and Conflict of Interest: DOE DE-SC0017057 NIH R44CA183390 NIH R01CA188300 NIH R43CA183390 NIH U19AI067769

Contact Email: