Program Information
Non-Local Means Denoising of SG-KS-4D-MRI Improves the Accuracy of Deformable Registration and Pancreas Tumor Segmentation
J Jin2 , E McKenzie1 , S Gou2 , G Yang2 , Z Fan1 , D Li1 , R Tuli1 , H Sandler1 , B Fraass1 ,W Yang1* (1) Cedars-Sinai Medical Center, Los Angeles, CA, (2) Xi'dian University, Xi An, Shan Xi
Presentations
TH-CD-204-8 (Thursday, July 16, 2015) 10:00 AM - 12:00 PM Room: 204
Purpose:Self-gating based K-space sorted 4D-MRI (SG-KS-4D-MRI) has been shown to be a promising method to generate high resolution respiration phase resolved 4D-MRI images. However, higher imaging noise was observed compared to breath-hold 3D-MRI. In this study, we apply a non-local means denoising filter on the pancreas SG-KS-4D-MRI images and then test its impact on the accuracy of deformable registration and automated pancreas tumor segmentation.
Methods: A non-local means denoising method, block matching 3D (BM3D), was adapted to search in the axial slices adjacent to the reference image patch for patches with high similarity. The located image patches were stacked to a 3D block based on similarity. 3D denoising was applied to the block in the wavelet domain, which was then followed by aggregation to reconstruct the denoised 4D-MRI image. The pancreas tumor was manually contoured in the first phase near the end of exhalation for both the raw and denoised 4D-MRI images with the same volume for each patient. B-spline deformable registration was applied to subsequent phases for contour propagation. The consistency of tumor volume, the center-of-mass trajectories, and the agreement between the center-of-mass locations between the first and last end-of-exhalation phases was compared using Wilcoxon-Rank test.
Results: BM3D visually reduced image noise and improved contrast without compromising soft tissue edge definitions. The propagated tumor volume was slightly (1%) but statistically significantly (p=0.008) larger in the denoised 4DMRI. The standard deviation in the tumor volumes was significantly (p=0.01) reduced by 25% with denoised 4D-MRI. The tumor center-of-mass trajectories generally become more regular showing more reproducible (statistically significant improvement in the anterior-posterior direction) in the two phases near the end-of-exhalation positions.
Conclusion: The non-local means denoising filter (BM3D) is shown to improve the noisy 4D-MRI image conspicuity that can be appreciated visually and quantitatively using deformable image registration.
Funding Support, Disclosures, and Conflict of Interest: This study is supported in part by NIH 1R03CA173273
Contact Email: