Model selection for high b-value diffusion-weighted MRI of the prostate

Publication date: February 2018
Source:Magnetic Resonance Imaging, Volume 46
Author(s): Yousef Mazaheri, Andreas M. Hötker, Amita Shukla-Dave, Oguz Akin, Hedvig Hricak
PurposeTo assess the abilities of the standard mono-exponential (ME), bi-exponential (BE), diffusion kurtosis (DK) and stretched exponential (SE) models to characterize diffusion signal in malignant and prostatic tissues and determine which of the four models best characterizes these tissues on a per-voxel basis.Materials and methodsThis institutional-review-board-approved, HIPAA-compliant, retrospective study included 55 patients (median age, 61years; range, 42–77years) with untreated, biopsy-proven PCa who underwent endorectal coil MRI at 3-Tesla, diffusion-weighted MRI acquired at eight b-values from 0 to 2000s/mm². Estimated parameters were apparent diffusion coefficent (ME model); diffusion coefficients for the fast (Dfast) and slow (Dslow) components and fraction of fast component, ffast (BE model); diffusion coefficient D, and kurtosis K (DK model); distributed diffusion coefficient DDC and α for (SE model). For one region-of-interest (ROI) in PZ and another in PCa in each patient, the corrected Akaike information criterion (AICc) and the Akaike weight (w) were calculated for each voxel.ResultsBased on AICc and w, all non-monoexponential models outperformed the ME model in PZ and PCa. The DK model in PZ and SE model in PCa ROIs best fit the greatest average percentages of voxels (39% and 43%, respectively) and had the highest mean w (35±16×10⁻² and 41±22×10⁻², respectively).ConclusionDK and SE models best fit DWI data in PZ and PCa, and non-ME models consistently outperformed the ME model. Voxel-wise mapping of the preferential model demonstrated that the vast majority of voxels in either tissue type were best fit with one of the non-monoexponential models. At the given SNR levels, the maximum b-value of 2000s/mm² is not sufficiently high to identify the preferred non-monoexponential model.



http://ift.tt/2z2Smw5