Comparison of Denoising Methods On Dynamic Contrast-Enhanced MRI
R He, Y Ding, K Wahid, A Mohamed, B McDonald*, M Naser, J Wang, K Hutcheson, C Fuller, S Lai, MD Anderson Cancer Center, Houston, TX
Presentations
PO-GePV-M-285 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To compare denoising methods on dynamic contrast-enhanced (DCE) MRI with a validated DCE algorithm, Standard Tofts-Kety model, Extended Tofts-Kety model, and Digital Reference Object (DRO).
Methods: Efficient denoising methods are critical for clinical applications. Towards this end, a previously validated DCE algorithm based on Julia language (DCEMRI.jl) is used as test-bed to compare denoising methods from two different categories to explore their ability in improving the recovery of DCE kinetic parameters from noisy data of DRO. The denoising algorithms we used are based on (1) Piece-wise Smoothness assumption such as Total Variation (TV) and simultaneously spatial and temporal High Order TVs (HOTVs), and (2) Self-Similarity such as Non-Local Means (NL-Means) and its adaptive version (ASCM). While DCEMRI.jl complete both linear and non-linear fitting algorithms for DCE kinetic models, and previously validated on DROs, we are especially interested in linear fitting method for its running speed and parametric map quality. By using DROs for comparing denoising methods, we can easily and precisely obtain the quantitative metrics of the recovered kinetic parametric maps by referring to the ground truth of kinetic parameters.
Results: All denoising methods can considerably improve the recovery of parametric maps. In Linearized Extended Tofts-Kety model, ASCM outperforms others in RMS error (RMS, from [35.7 41.5 50.12] to [8.6 22.9 26.86] %) and the concordance correlation coefficients (CCC, form [0.97 0.69 0.97] to [0.998 0.844 0.997] %) for Ktrans, Ve, and Vp. In Linearized Standard Tofts-Kety model, HOTVs outperforms others in RMS (from [55.6 56.62] to [28.26 12.77] %) and CCC (from [0.67 0.80] to [0.82 0.994] %) for Ktrans and Ve. In computational speed, NL-Means is the fastest, and ASCM and HOTVs are similar.
Conclusion: The denoising methods compared in this study can efficiently improve the recovery of DCE kinetic parametric map for noisy MRI acquisitions.
Keywords
Not Applicable / None Entered.
Taxonomy
IM/TH- Image Analysis (Single Modality or Multi-Modality): Image processing
Contact Email