Offline Reconstruction of Diffusion MRI Acquisitions for Comparison Between Complex PCA-based and AI-based Denoising

Purpose: Optimal diffusion MRI (dMRI) reconstruction for image denoising is often unavailable from scanner reconstruction. In this work we make available an offline reconstruction pipeline for GE dMRI acquisitions, giving access to complex, unfiltered dMRI data. Furthermore, we compare the efficacy of GE HealthCare's AIR-Recon DLTM (ARDL), a proprietary convolutional neural network-based reconstruction and denoising approach, to open-source patch-based MPPCA and NORDIC denoising methods on high-resolution dMRI data. Methods: We developed an end-to-end offline dMRI reconstruction pipeline for GE HealthCare acquisitions, building on and augmenting the Orchestra software development kit and validated its output against scanner reconstruction. We used it to compare MPPCA, NORDIC and ARDL denoising approaches, considering underlying metrics reflecting noise variance and noise floor suppression, such as the signal dynamic range and ADC in highly anisotropic areas, noise variance in maximally-suppressed dMRI signal and downstream measurements, such as fibre orientation estimation and white matter tractography. Results: Our validated offline reconstruction starts from single-channel complex k-space data and allows channel combination, support of various in-plane/out-of-plane accelerations and partial Fourier reconstruction methods and retrospectively switching filters off. Unlike scanner reconstruction, our pipeline provides access to complex dMRI data, allowing denoising in the complex domain, which showed superior noise floor suppression compared to magnitude constrained denoising. Comparisons across denoising methods suggest improved spatial resolution, contrast-to-noise and more robust fibre orientation estimation when using patch-based approaches compared to ARDL. Conclusion: We found significant gains in dMRI data quality when using the proposed offline reconstruction pipeline, allowing denoising to occur in the complex domain, both for reducing noise-induced variance and bias. MPPCA and NORDIC (4D patch based) outperformed ARDL (2D) in terms of spatial resolution, reduction of noise-floor bias and variance.