Evaluation of Denoising Methods for ADC Accuracy on Low-Field MR-Linac Diffusion-Weighted Imaging
Abstract
Purpose
We compared denoising methods for Diffusion-Weighted Imaging (DWI) of phantoms acquired on a low-field MR-Linac to determine which method best preserves Apparent Diffusion Coefficient (ADC) accuracy and image feature detection. Our goal was to quantify image improvement for future denoising implementation on thoracic and pelvic DWI, which often has a low signal-to-noise ratio.
Methods
A QAlibreMD and Magphan phantoms were imaged on a 0.35T MR-Linac using 1 and 8 averages, respectively, and 16 averages. Low average DW images were denoised with Total Variation (TV) and Wavelet from scikit-image, and Non-Local Means (NLM), and Marcenko-Pastur PCA (MPPCA) from DIPY. ADC maps were calculated using a mono-exponential fit. Mean ADC values were measured in 13 regions of interest (vials with varying PVP concentration) on the diffusion phantom and compared to scanner ADC maps acquired with 16 averages. MagPhan mimicked the size and shape of a human torso, and provided measurements of the number of spheres detected, FWHM of the Point Spread Function (PSF), and 10-90% Edge Spread Function (ESF) values.
Results
For QAlibre vials, mean ADC values were preserved within 2% of the 16 average reference for all methods except TV, which underestimated ADC by 43-88%. NLM showed the best ADC conservation. TV and NLM produced the lowest standard deviation, corresponding to the greatest noise reduction. Sphere detection in the MagPhan ADC maps increased with TV and MPPCA (399 and 389, respectively), while NLM detected only 365 spheres. The ESF and PSF, which quantify image clarity in terms of edge sharpness and point blurriness, improved most with MPPCA. TV and wavelet methods exhibited little improvement.
Conclusion
Phantom testing revealed that MPPCA performed best in yielding accurate ADC maps and improving feature detectability, with NLM demonstrating comparable results. TV and wavelet were discarded due to inconsistent ADC and high noise, respectively.