Program Information
Low-Dose CT Image Denoising Using An Optimized Wiener Filter in the BM3D Algorithm
T Zhao*, J Hoffman , M McNitt-Gray , D Ruan , UCLA School of Medicine, Los Angeles, CA
Presentations
WE-F-605-4 (Wednesday, August 2, 2017) 1:45 PM - 3:45 PM Room: 605
Purpose: The block-matching 3D (BM3D) algorithm provides state-of-the-art denoising performance when images are corrupted with additive white Gaussian noise (AWGN). In low-dose CT imaging, noise is object-dependent and not AWGN. This study investigates the denoising of low-dose CT images using BM3D with a modified Wiener filter that is adapted to the power spectral properties of the image noise.
Methods: In the original BM3D algorithm, the Wiener filtering is applied in the transform domain to a post-threshold signal for enhanced denoising. We first evaluated the performance of the current Wiener filtering on low-dose CT images, and analyzed the specific noise properties. We then derived the optimal coefficients of Wiener filter based on the minimum mean-square-error (MMSE) criterion, using the power spectral density of noise and the correlation between signal and noise. Performance comparison with the original BM3D method was performed on thoracic CT image datasets, using paired full-dose and simulated 10%-dose images.
Results: The low-dose CT noise, generated by a pipeline with stochastic noise generation in the projection domain, presented distinct non-Gaussian characteristics and was highly correlated with image intensity. Evaluation showed that the current Wiener filter in BM3D algorithm yielded little denoising enhancement in low-dose CT images. In contrast, the proposed Wiener filtering strategy achieved improvement with (.93, .95)dB performance gain in mean and median peak signal-to-noise ratio (PSNR). Paired t-test of the PSNR between denoising using original and proposed Wiener filter yielded p-value of 3.83E-8.
Conclusion: Tailoring the Wiener filter in BM3D algorithm to problem statistics can improve denoising performance. The benefit of accounting for signal/noise correlation in low-dose CT images is demonstrated in this work.
Funding Support, Disclosures, and Conflict of Interest: Funding support for this research was provided in part by the University of California Office of the President Tobacco-Related Disease Research Program (UCOP-TRDRP grant #22RT-0131) and the National Cancer Institute's Quantitative Imaging Network (QIN grant U01-CA181156).
Contact Email: