Intro

Comparison of ROI CT reconstruction methods

Overview of the methods being compared (see tables below):

Method name	Description	ROI CT method	Sparsity/regularity domain
SIRA-FIDEL	Sparsity induced reconstruction algorithm with maximal data fidelity (and L1 prior, haar wavelet, one scale)	Yes	Ridgelet
SIRA	Sparsity-induced reconstruction algorithm with fidelity and consistency trade-offs (and L1 prior, haar wavelet, one scale)	Yes	Ridgelet
SIRA-L2	Regularity-induced reconstruction algorithm with maximal data fidelity (using L2 prior, haar wavelet, one scale)	Yes	Ridgelet
OSSART-TV	Ordered Subset Simultaneous Algebraic Reconstruction Technique, ROI CT version with Total Variation Regularization	Yes	TV (image domain)
OSSART-shearlet	Ordered Subset Simultaneous Algebraic Reconstruction Technique, ROI CT version with L1 Shearlet Regularization	Yes	shearlet (image domain)
CS-TV	Classical compressed sensing method with TV regularization	Yes	TV (image domain)
CS-shearlet	Classical compressed sensing method with L1 shearlet regularization	Yes	shearlet (image domain)
DBP	Differentiated backprojection method	Yes	N/A
MLEM	Maximum likelihood expectation maximization algorithm	No	N/A
LSCG	Least-squares conjugate gradient method	Yes	N/A

Info:

The PSNR values are calculated with respect to a full view LSCG reconstruction (i.e. without truncated projection data). Note that this penalizes methods that perform regularization in the image domain/ridgelet domain (e.g. to denoise the images). Moreover, the PSNR is calculated over a disk with the exact ROI radius. Interpret the PSNR metric therefore with care.
Two real datasets are being processed (mouse and abdomen). These datasets are acquired with a micro CT scanner from Ghent University hospital.
Note: For some reason, the webbrowsers do not display the images as they are (they perform some processing mostly contrast/brightness enhancement). However, for all images the processing is the same.
Click on the table headers for a comparison of different methods/ROI radii.

Quick summary of the results:

The SIRA methods significantly outperform all other methods in terms of PSNR
A reliable reconstruction is about >20dB. It can be seen that all methods (except SIRA) do not work reliably for ROI radius < 48
Even for small ROI radii, SIRA-FIDEL offers a high reconstruction quality.
For larger ROI radii, the contrast of the reconstructed images of OSSART seems to be not right. I am having a look at this.
The non-SIRA methods mostly suffer from artifacts near the boundary of the ROI (bright cupping artifact). In some cases, they destroy features inside the ROI.