Premium
SU‐C‐206‐04: Region‐Specific Total‐Variation Regularization for X‐Ray CT Reconstruction
Author(s) -
Xu Q,
Han H,
Xing L
Publication year - 2016
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4955586
Subject(s) - regularization (linguistics) , pixel , image noise , weighting , mathematics , algorithm , cluster analysis , iterative reconstruction , total variation denoising , computer science , gradient noise , noise (video) , value noise , artificial intelligence , pattern recognition (psychology) , mathematical optimization , noise measurement , noise reduction , image (mathematics) , physics , noise floor , acoustics
Purpose: Total‐variation (TV) regularization methods are very effective to handle the non‐sufficient data reconstruction, such as few‐view problem, low‐dose data, etc. Usually one single regularization parameter is employed to balance the data fidelity and the TV minimization penalty in all regions of the object. However, noise and structural details vary from region to region. Here we proposed a method to select the regularization parameter adaptively by the region specific noise. Methods: The image was reconstructed by minimizing the data fidelity term and the TV term alternatively. After the image was updated by the data fidelity step, its noise distribution was estimated out by calculating the variance of neighboring pixels. Then a fuzzy clustering method was applied on it to proportionally assign the membership of each pixel into several clustering groups which correspond to regions with different noise levels, respectively. A heuristic function was developed to describe the relationship between the noise and the regularization parameter. The TV was performed under different parameters which were determined by the noise levels. The final image of TV minimization step was generated by fusing these results according to the weighting coefficients determined by the fuzzy membership. After that a new iteration was repeated until the stopping criterion was satisfied. A view down sampled human thorax dataset was used to evaluate the proposed method. Results: Since the noise distribution was not uniform, the conventional TV method with one single parameter cannot make a good balance between the noise and structural details in all regions simultaneously while the proposed region‐specific method can achieve a better result. The structure similarity index (SSIM) increased 2.1% and the root mean squared error (RMSE) decreased 9% in the vertebra region with comparable results in other regions. Conclusion: The proposed region specific regularization can improve the current TV‐based reconstruction.