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

Offline Reconstruction of Diffusion MRI Acquisitions for Comparison Between Complex PCA-based and AI-based Denoising | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Offline Reconstruction of Diffusion MRI Acquisitions for Comparison Between Complex PCA-based and AI-based Denoising View ORCID Profile Francesco D’Antonio , View ORCID Profile Shaun Warrington , View ORCID Profile Jose-Pedro Manzano-Patron , Paul S. Morgan , View ORCID Profile Stamatios N. Sotiropoulos doi: https://doi.org/10.1101/2025.11.07.682283 Francesco D’Antonio 1 Sir Peter Mansfield Imaging Centre, School of Medicine, University of Nottingham , United Kingdom 2 Mental Health and Clinical Neurosciences, School of Medicine, University of Nottingham , United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Francesco D’Antonio Shaun Warrington 1 Sir Peter Mansfield Imaging Centre, School of Medicine, University of Nottingham , United Kingdom 2 Mental Health and Clinical Neurosciences, School of Medicine, University of Nottingham , United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Shaun Warrington Jose-Pedro Manzano-Patron 1 Sir Peter Mansfield Imaging Centre, School of Medicine, University of Nottingham , United Kingdom 2 Mental Health and Clinical Neurosciences, School of Medicine, University of Nottingham , United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Jose-Pedro Manzano-Patron Paul S. Morgan 1 Sir Peter Mansfield Imaging Centre, School of Medicine, University of Nottingham , United Kingdom 2 Mental Health and Clinical Neurosciences, School of Medicine, University of Nottingham , United Kingdom 3 National Institute of Health & Care Research Nottingham Biomedical Research Centre, Nottingham University Hospitals , United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site Stamatios N. Sotiropoulos 1 Sir Peter Mansfield Imaging Centre, School of Medicine, University of Nottingham , United Kingdom 2 Mental Health and Clinical Neurosciences, School of Medicine, University of Nottingham , United Kingdom 3 National Institute of Health & Care Research Nottingham Biomedical Research Centre, Nottingham University Hospitals , United Kingdom Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Stamatios N. Sotiropoulos For correspondence: stamatios.sotiropoulos{at}nottingham.ac.uk Abstract Full Text Info/History Metrics Supplementary material Preview PDF Abstract Purpose Optimal diffusion MRI (dMRI) data 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 dMRI data. Furthermore, we compare the efficacy of GE HealthCare’s AIR-Recon DL™ (ARDL), a proprietary convolutional neural network-based reconstruction and denoising approach, to open-source PCA-based MPPCA SVS and NORDIC denoising methods on high-resolution dMRI data. Methods We developed an end-to-end offline dMRI reconstruction pipeline for GE HealthCare acquisitions, augmenting the Orchestra software development kit, and validated its output against scanner reconstruction. We used it to compare MPPCA SVS , NORDIC and ARDL denoising approaches, considering underlying metrics reflecting noise variance and bias, such as the ADC profiles in highly anisotropic areas, and downstream measurements, such as fiber orientation estimation and white matter tractography. Results Our validated offline reconstruction supports various in-plane/out-of-plane accelerations and partial Fourier reconstruction methods. Unlike scanner reconstruction, it provides access to complex dMRI data, enabling denoising in the complex domain, which demonstrated superior noise floor suppression compared to magnitude-constrained denoising. PCA-based denoising methods had improved spatial resolution, contrast-to-noise and more robust fiber orientation estimation compared to ARDL. Conclusion We found significant gains in dMRI data quality when using the proposed offline reconstruction pipeline, allowing complex-domain denoising to obtain high-quality data at high spatial resolution and b-value, using a wide-bore scanner and a standard PGSE EPI sequence. MPPCA SVS and NORDIC (4D PCA-based) outperformed ARDL (2D) in terms of spatial resolution, reduction of noise-floor bias and variance. 1. Introduction The value of denoising diffusion MRI (dMRI) data has been demonstrated in various scenarios, enhancing fidelity of dMRI-derived microstructural estimates [ 1 – 3 ], increasing spatial resolution [ 2 , 4 ], and improving consistency between scanners [ 5 ]. DMRI contrast is derived from signal attenuation, leading to low SNR, especially when increasing resolution or diffusion weighting. This results in low signals being rectified to the noise floor [ 6 ], leading to biases in parametric maps and downstream fiber modelling [ 3 , 7 – 9 ]. To overcome these challenges, several denoising approaches have been developed using PCA-based, low-rank reduction of signals [ 5 , 10 – 17 ]. These methods exploit the spatial, angular and diffusion-weighted redundancy in dMRI signals to estimate and remove noise-dominant principal components from local patches. A breakthrough in the field occurred when Random Matrix Theory was used to derive data-driven estimates for the threshold between noise and signal-dominant principal components [ 12 , 13 ]. This is achieved by fitting the Marchenko-Pastur (MP) distribution [ 18 ] to the distribution of principal components, eliminating the need for a separate estimate of the noise variance [ 14 , 19 ]. Refinements for threshold estimation, which account for signal-dominant components that alter the theoretical MP distribution, have been implemented to better estimate the signal-noise boundary [ 10 , 16 , 20 ]. Additionally, to estimate noise-free signal more accurately, optimal shrinkage of principal components is used to mitigate eigenvalue inflation due to noise [ 21 ] using nuclear [ 1 ] or Frobenius norms [ 5 , 16 ]. Regardless of the exact implementation, it has been demonstrated that the domain (magnitude/complex) into which dMRI denoising is applied can have significant implications for subsequent analysis, affecting noise-induced variance, bias, and true resolution in different ways [ 1 , 2 , 5 , 22 ]. A challenge, however, is that across scanner vendors it is not straightforward (or not feasible at all in some cases) to control reconstruction parameters, such as apodization filters, or even access complex data, which would improve denoising outcomes. In addition to these PCA-based approaches, scanner manufacturers have recently introduced AI-driven reconstruction and denoising solutions to enhance resolution and SNR while maintaining or reducing scan time [ 23 – 27 ]. These methods are integrated into the scanner reconstruction process, making them easily implementable in clinical protocols, and benefit from having access to unfiltered, complex domain data for optimal noise reduction. One such method is AIR™ Recon DL (ARDL) from GE HealthCare (Milwaukee, USA), a reconstruction and denoising convolutional neural network trained on millions of MRI images from various scan locations and acquisition methods, which is used to and remove thermal noise and Gibbs ringing artefacts [ 23 ]. ARDL operates on 2-dimensional image slices in the complex domain. The trade-off between resolution, noise, and scan time typically governs acquisition choices for MRI, particularly for noise-prone modalities, such as dMRI, making it an ideal modality for assessing the performance of this denoising method. However, it is unclear how DL-based approaches, such as ARDL, perform when compared against PCA-based approaches. In this paper, we address the above questions. Firstly, we develop an end-to-end offline reconstruction pipeline compatible with GE dMRI acquisitions. It takes raw, single-channel, complex k-space data as input and provides access to complex dMRI data. The pipeline supports retrospective switching of filters and partial Fourier reconstruction methods and is compatible with in-plane and out-of-plane accelerated acquisitions. This is necessary as scanner reconstruction on GE scanners does not provide access to complex dMRI data. Secondly, we demonstrate the application of this offline reconstruction pipeline by comparing PCA-based denoising in the complex domain with GE’s proprietary ARDL on standard PGSE acquisition from a wide-bore scanner. Specifically, we begin by validating the pipeline by comparing the scanner’s and the pipeline’s magnitude reconstruction across different acceleration and reconstruction parameters. We then compare ARDL with two commonly used denoising methods, NORDIC [ 11 ] and MPPCA SVS [ 5 , 12 , 13 ] (MPPCA with singular value shrinkage), as they can be easily implemented in the complex domain. We also include MPPCA SVS denoising in the magnitude domain to highlight improvements in denoising outcomes when using the offline reconstruction pipeline’s complex-domain data, over magnitude-constrained denoising available from native scanner reconstruction. To evaluate denoising outcomes, we consider quantitative metrics of noise variance, noise floor bias and proxies of noise distribution within the brain. We also evaluate the downstream effects of denoising through fiber orientation estimation and white matter bundle segmentation with probabilistic tractography. 2. Methods 2.1 Image Reconstruction Pipeline We developed a comprehensive and scalable offline reconstruction pipeline for dMRI acquisitions on GE HealthCare scanners, starting from single-channel k-space data from individual coil elements and producing coil-combined images in both complex and magnitude domains. The pipeline can modify several reconstruction steps and handle multiple accelerated acquisition options. We used Python (version 3.12.6) and built upon the C++ Orchestra software development kit (SDK), which provides simple example reconstruction scripts and is available through GE HealthCare’s WeConnect website ( https://weconnect.gehealthcare.com/ ) with the appropriate research license. The pipeline takes as input ScanArchive.h5 files (k-space, individual coil data) and outputs DICOMs and NIfTI images (either magnitude or complex) as shown in Figure 1 . Complementing further the capabilities of the default Orchestra SDK scripts, our workflow allows: Download figure Open in new tab Fig. 1. Outline of the offline reconstruction pipeline. Orange steps are modified C++ Orchestra SDK code; in blue are additional steps using custom Python code. Boxes with only colored borders (filled in white) are optional steps. The pipeline is implemented within a container for reproducibility and ease of use across operating systems. Reconstruct channel-combined complex (magnitude/phase or real-rotated) data. Support of in-plane (ASSET-SENSE and ARC-GRAPPA) and out-of-plane accelerations, with proper reconstruction of HyperBand acquisitions for simultaneous multi-slice (SMS) accelerations. Retrospective reconstruction of partial Fourier acquisitions, using zero-filling, POCS [ 28 ] and Homodyne [ 29 ] methods. Access to single-channel complex images and extraction of coil sensitivity profiles to be used for sensitivity-weighted channel combination. Proper indexing of dMRI volumes and slices for all combinations of in-plane (SENSE-like, GRAPPA-like) and out-of-plane (SMS) accelerations. Optional flags to modify several reconstruction steps retrospectively, including switching on/off gradient non-linearity correction, on/off apodization filters, on/off high-order eddy current correction and switching between complex (zero-filling) and homodyne reconstruction [ 29 ] for partial Fourier acquisitions. Accounting for and preserving phase encoding direction and correct creation of diffusion b values (bval) and diffusion-weighted direction vector (bvec) files for subsequent spin-echo fieldmap-based corrections of susceptibility-induced distortions. To enable the reproducible and straightforward implementation of the pipeline across various operating systems and high-performance clusters, the offline reconstruction is embedded into Docker ( https://docs.docker.com/ ) or Singularity/Apptainer ( https://apptainer.org/ ) containers. Additionally, the C++ source code for the pipeline is available for customisation. The pipeline and source code is available through GE HealthCare’s WeConnect website with the appropriate research license. To evaluate the capabilities of the pipeline, data using SENSE-like acceleration (ASSET factors 1 and 2), GRAPPA-like acceleration [ 30 ] (ARC factors 1 and 2), simulataneous multi-slice [ 31 – 33 ] (HyperBand factors 1 and 4), and partial Fourier reconstruction methods (homodyne or zero-filled), were acquired and reconstructed offline, using the developed reconstruction pipeline, as well as on the scanner. The magnitude reconstruction from both approaches was compared to validate the offline pipeline’s output. 2.2 Denoising Comparisons The offline reconstruction was developed with denoising in mind and hence offers the ability to perform denoising either in magnitude or complex domain at various stages of the reconstruction, for instance, prior to or after channel combination. In this work, we considered two PCA-based denoising methods applied after channel combination: MPPCA SVS : using 7×7×7 patches (using the smallest odd N such that NxNxN > number of volumes acquired, using the heuristic in MRTrix [ 34 ]), using eigenvalue threshold estimation from Cordero-Grande et al. [ 10 ], and singular value shrinkage using Frobenius norm [ 21 ], as implemented in the DESIGNER pipeline (v2.0.11) [ 35 , 36 ]. This was applied to both complex (real rotated homodyne [ 29 ]) data (MPPCA* SVS ) and magnitude (|MPPCA| SVS ) data. NORDIC: using 13×13×13 patches (so that the number of voxels in a patch and the number of diffusion volumes are approximately in a ratio of 11:1, as suggested in the NORDIC code) ( https://github.com/SteenMoeller/NORDIC_Raw ) [ 11 ] applied to complex (real rotated homodyne) data (NORDIC*). We did not provide noise only volumes, therefore the NORDIC method normalizes the data using an estimated g-factor map using 2D MPPCA [ 12 , 13 ] noise residuals, then calculates low rank thresholds using generated patch-seized noise matrices. We subsequently used the reconstruction pipeline to compare these PCA-based denoising approaches with GE’s AI-based ARDL. Starting from raw k-space data, we retained the real-channel rotated data from homodyne reconstruction [ 29 ] (referred to as complex data in the rest of the paper) and used this for denoising in the complex domain. ARDL, which also operates in the complex domain, was applied with the “high” denoising weighting hyperparameter (from low, medium, and high options from the scanner console interface) on the same data, using retrospective reconstruction on the scanner. As ARDL denoising requires a fixed-size 256×256 input matrix by design, k-space was manipulated by placing the acquired k-space in a zero-filled 256×256 matrix during reconstruction. After ARDL was applied, data were downsampled to the acquisition matrix size using spline interpolation [ 37 , 38 ]. An overview of the reconstruction and processing pipeline is shown in Figure 2 . Download figure Open in new tab Fig. 2. Overview of processing steps for comparison across denoising methods. For ARDL reconstruction and denoising, the k-space is resampled to a 256×256 matrix and later (post-ARDL) downsampled to the native matrix size. 2.2 Data Acquisition Data for offline reconstruction pipeline validation were acquired on a GE HealthCare 3T Premier wide-bore scanner (Software Version MR29.2; GE HealthCare, Milwaukee, USA) at the University of Nottingham. Data from a single participant was acquired using 2 mm isotropic, 6 directions, b=1000 s/mm 2 PGSE-EPI at various acceleration parameters: no acceleration (TR/TE = 8.0s/57.2ms), ASSET=2 (TR/TE = 8.0s/57.3ms), ARC=2 HyperBand=4 (TR/TE = 8.0s/59.7ms) as well as a high-resolution 1.3 mm isotropic HCP-style data [ 39 , 40 ], (98 volumes/shell, two shells: b=1500, 3000 s/mm 2 , HyperBand=4, TR/TE=4.8s/88.3ms). Data for denoising comparisons were acquired at the Cambridge University Hospitals NHS Foundation Trust (Cambridge, UK) for four healthy participants (two female, age range 27 to 54) on a GE HealthCare 3T Premier wide-bore scanner with a 48-channel head coil (Software Version MR30.1; GE HealthCare, Milwaukee, USA). Data were acquired using a PGSE-EPI sequence following the acquisition protocols used in the Human Connectome Project (HCP) [ 39 , 40 ], with 1.25 mm isotropic voxel size, 98 volumes/shell, two shells: b=1500, 3000 s/mm 2 , and 14 b=0 s/mm 2 volumes throughout the acquisition with one reverse gradient (AP/PA) b=0 s/mm 2 acquired prior to the main data, HyperBand=4, TR/TE=4.8s/88.3ms, 60% partial Fourier with Homodyne reconstruction. A 1 mm isotropic T1-weighted MPRAGE with higher order shimming was acquired to create white matter masks and improve registration to MNI templates. All data were acquired after informed consent was obtained from all participants. Ethical approval was respectively obtained from the University of Nottingham Faculty of Medicine and Health Sciences Research Ethics Committee (B12012012a_073) and from the East of England regional review board, Cambridgeshire and Hertfordshire Research Ethics Committee (08-H0311-117). 2.3 Pre-Processing and Analysis Post-denoising, identical distortion correction across denoising methods was carried out using a standard FSL-eddy pipeline (FSL v6.0.6.4) [ 41 , 42 ] on the images, including susceptibility-induced, eddy-current distortion and motion correction. We evaluated several aspects of denoising impact, from raw data to downstream analysis, as recommended in Manzano-Patron et al. [ 2 ]. To assess denoising efficacy with respect to variance suppression, SNR and angular CNR were obtained using FSL-eddyQC [ 43 ]. To assess efficacy with respect to bias and noise floor suppression, we examined the distribution, dynamic range and ADC profile of the signal, before and after denoising. To account for noise spatial non-stationarities, noise distributions were estimated from inside the brain (rather than background) using maximally attenuated (i.e. high b-value) CSF signal from the ventricles [ 44 ]. Noise-floor-induced signal rectification was assessed by examining the signal attenuation and ADC measurements of parallel vs perpendicular (i.e. maximally vs minimally attenuated) to principal fiber directions (DTI V 1 ) in highly anisotropic voxels [ 44 ], in the corpus callosum, using single-subject warped masks. Downstream effects of denoising methods were also assessed through fiber orientation estimation and probabilistic tractography, carried out using FSL’s bedpostX [ 45 , 46 ] and XTRACT [ 47 ] for each denoising mode and we compared session-wise tract density maps with average probabilistic tract density maps from a 50-subject HCP average [ 47 ]. 3. Results 3.1 Offline Reconstruction Features and Validation We validated the offline reconstruction by comparing magnitude images between the scanner and the offline reconstruction pipeline, as complex dMRI data are not available from the scanner reconstruction. We performed comparisons across resolutions and b-values, as shown in Figure 3 , for various in-plane (no acceleration, ASSET=2, and ARC=2) and out-of-plane (no HyperBand and HyperBand=4) acceleration. In all cases, the offline magnitude reconstruction matched the scanner reconstruction almost perfectly when no SMS acceleration was used. When SMS acceleration was used minor signal differences were seen within the brain, but these differences were below 1% in absolute value. Download figure Open in new tab Fig. 3. Comparison of offline magnitude and scanner reconstruction. Offline pipeline reconstruction and voxel-wise percentage difference in signal intensity between offline and scanner reconstruction for (top left to bottom right) 2mm isotropic no acceleration, ASSET = 2, ARC=2 + HyperBand=4 and 1.3mm isotropic HyperBand= 4 data. Minimal differences are observed when no acceleration or ASSET is used, and differences are below 1% in absolute value when HyperBand or ARC is used. Figure 4a highlights some modifications made to the C++ Orchestra SDK to enable correct indexing of slices, correctly account for the reference scans in SMS acquisitions and correctly identify the phase encoding direction for blip-reversed acquisitions. We showcase examples of reconstructed complex images in Figure 4b , showing magnitude and phase reconstruction for a zero-filled acquisition and real-rotated homodyne acquisition (exemplified by negative values in the CSF signal). The offline pipeline also gives access to complex single-channel images and respective coil sensitivity profiles, as shown in Figure 4c , which can be retrospectively reconstructed using custom homodyne or POCS algorithms. Download figure Open in new tab Fig. 4. Offline reconstruction modifications and additional features. a) Images highlighting some modifications to the original Orchestra SDK reconstruction, from top to bottom: correcting slice indexing, SMS reconstruction and phase encoding direction reconstruction. b) Examples of pipeline complex zero-filled (magnitude and phase) or homodyne (real rotated) reconstruction, not available from scanner reconstruction. c) Example of single channel reconstruction showing single channel magnitude, phase and sensitivity profiles for channel combination. 3.2 Denoising Comparisons The effect of denoising on exemplary slices for b=3000 s/mm 2 (top) and b=1500 s/mm 2 (bottom) shells is shown in Figure 5 . To showcase the increased ability to discriminate anatomy post-denoising, sulci between the insula and adjacent lobes are highlighted. The location and size of the sulci is compared between high signal CSF in b=0 s/mm 2 slices and low signal in diffusion weighted slices. All denoising approaches substantially improve the images over the raw data. MPPCA SVS denoised diffusion slices show the highest anatomical correspondence with the b=0 s/mm 2 slice. Noise floor-induced bias in magnitude denoised data (|MPPCA| SVS ) is highly visible when compared to complex denoised data (MPPCA* SVS , NORDIC or ARDL), highlighting the advantages of denoising complex, pipeline-reconstructed data over magnitude-denoised data. Furthermore, as highlighted in the figure inlays, anatomical details are more visible when applying MPPCA* SVS or NORDIC compared to ARDL, with MPPCA* SVS showing qualitatively the best performance across all approaches. Download figure Open in new tab Fig. 5. Comparison of raw and denoised data for b=3000 s/mm 2 (top) and b=1500 s/mm 2 (bottom) slices. A raw b=0 s/mm 2 slice is shown on the left to provide anatomical benchmarks for diffusion weighted volumes. Different columns correspond to raw data, magnitude-denoised |MPPCA| SVS , complex-denoised MPPCA* SVS , complex-denoised NORDIC* and complex-denoised ARDL with high denoising weighting. Smaller panels highlight sulci (highly visible in b=0 s/mm 2 raw data) to compare improvements in resolution and anatomical fidelity of denoised data. We subsequently performed quantitative assessments. Figures 6 – 8 provide an overview of metrics related to changes in variance and bias of the signal due to denoising. Estimates of SNR across the b=0 s /mm 2 volumes and angular CNR per shell are presented in Figure 6 , showing better performance for 4D PCA-based denoising methods compared to 2D ARDL. Download figure Open in new tab Fig. 6. SNR (across b=0 s/mm 2 volumes) and angular CNR estimates using FSL-EddyQC, thinner grey box plots represent single-subject data. Figure 7a presents the dMRI signal attenuation at a given b value, having the signal reordered based on the dot product of the corresponding diffusion-sensitisation gradient direction and the principal DTI V 1 in each voxel (i.e. parallel to perpendicular the main fiber orientation) and then averaged across a region of very anisotropic voxels (depicted by the mask in the body of the corpus callosum). We can observe how the signal is rectified in the magnitude domain and how all complex-domain denoising methods significantly increase the signal dynamic range and reduce the noise floor bias. The most significant decreases are associated with NORDIC and ARDL. In Figure 7b , the ADC peanuts present a similar picture with magnitude-domain denoising retaining a “squashed” peanut shape [ 7 ], which is recovered by all the complex denoising approaches. Download figure Open in new tab Fig. 7. Signal dynamic range and ADC measurements in the corpus callosum obtained using a registered mask and an FA>0.6 mask from RAW data (panel highlights example high FA voxel mask in red). a) Signal attenuation is measured by binning b=3000 s/mm 2 shell signals with respect to angle to DTI v1 direction and normalizing by the highest signal bin for each denoising mode. b) The ADC peanuts use the binned ADC measurement with respect to angle to DTI v1 direction for either the b=1500 s/mm 2 or b=3000 s/mm 2 shell. Finally, Figure 8 presents the distribution of maximally attenuated CSF signal in the ventricles, a proxy for the noise distribution, again showing increased noise floor suppression when denoising in the complex domain, with ARDL having the best performance. Taken together, these results suggest that denoising in the complex domain reduces noise-induced bias and variance, with ARDL performing the best in the removal of bias in the ventricles, NORDIC performing best in the removal of bias in regions of high diffusion anisotropy, and MPPCA* SVS performing the best in the reduction of variance. Download figure Open in new tab Fig. 8. Distribution of signal in the ventricles as a proxy for the noise distribution. Inlay highlights example ventricle mask (red) for a single subject. Downstream effects of denoising were assessed by exploring fiber orientation estimation and white matter bundle reconstructions, as seen in Figure 9 . Fiber orientation estimations using the multi-shell ball & sticks model (up to 3 orientations per voxel) are shown in Figure 9a , highlighting estimates around the centrum semiovale, as a location of complex (up to three-way) fiber crossings. Both MPPCA* SVS and ARDL denoised data show increased identification of two and three-way crossings, depicting orientations from the corpus callosum, the corticospinal tract and the longitudinal fasciculi running in and out of the coronal plane. These are mostly absent in the raw data. Also, MPPCA* SVS estimates are more qualitatively spatially continuous than ARDL, implying better overall performance. It is important to note that FSL’s BedpostX uses an automatic relevance detection (ARD) algorithm [ 48 ] to select the number of crossing fibers in a voxel, using a data-driven approach to select between one and three fibers to model each voxel’s signal. Download figure Open in new tab Fig. 9. a) Comparison of fiber orientation estimation, using FSL’s BedpostX, highlighting regions of complex fiber architecture involving multiple crossing fibers. Sticks are length-scaled by their corresponding volume fraction. Voxel shade represents the sum of anisotropic volume fractions. b) Results from probabilistic tractography (average of 4 subjects) using FSL’s XTRACT showing maximum intensity projections of superior longitudinal fasciculi (left) and correspondence with a 50-subject HCP average for all 42 tracts (right). Path distributions were thresholded at 0.1% and correlated against the corresponding path distributions of the HCP average. Fiber orientation estimates were subsequently used to perform probabilistic tractography and reconstruct 42 major white matter fiber bundles using FSL’s XTRACT. Figure 9b (left panel) examples participant-averaged tract density maps, highlighting results for the superior longitudinal fasciculi (SLF1, SLF2 and SLF3) and 50-subject HCP averages for comparison. We selected the SLFs as they pass through regions of complex fiber architecture. In the RAW data, tracking of the SLF1 was unsuccessful, whereas MPPCA* SVS and ARDL successfully tracked the SLF1, though ARDL incorrectly fit streamlines which more likely belong to SLF2. Other smaller differences can be observed, for instance MPPCA* SVS reconstructed SLF3 projects more anteriorly than the RAW or ARDL reconstruction. To quantify the differences further, violin plots on the right show the Pearson’s correlation between all 42 tract density maps of raw or denoised data and those of the HCP average, used as a reference. Denoising both with MPPCA* SVS and ARDL significantly improved correspondence to the reference, with MPPCA* SVS yielding better agreement overall. 4. Discussion In this work, we presented a reconstruction pipeline for GE dMRI acquisition, which uses single-channel k-space data to give access to unfiltered, channel-combined complex data. The pipeline also gives the freedom to alter reconstruction steps like apodization filters, partial Fourier reconstruction method (POCS/Homodyne/Zero-filled), and gradient warp fields. In line with previous finding which used data from other manufacturers [ 1 , 2 , 5 , 22 ], we demonstrate the benefits of denoising in the complex domain, using data obtained with the proposed pipeline, over magnitude-constrained denoising. This is particularly evident in reducing the effect of the noise floor bias, as observed in the signal dynamic range, ADC profiles in the corpus callosum, and noise distribution in the ventricles. Our offline reconstruction allowed us to obtain high-quality data at high spatial resolution and b value (1.3mm isotropic, b values up to 3000 s/mm 2 ), similar to HCP dMRI [ 39 ], even if using a wide-bore scanner, such as the GE Premier, and a standard PGSE EPI sequence. Previous work reports that denoising dMRI data reduces the variability of downstream metrics across scanners and protocols, with the greatest benefit when denoising in the complex domain [ 5 ]. So far, these comparisons have been limited to a single manufacturer across denoising domains [ 5 ], or using multiple manufacturers but constrained to magnitude domain denoising, with little control of the reconstruction algorithms used across vendors [ 49 ]. Our introduction of a flexible offline reconstruction pipeline for GE dMRI data allows to bridge the gap and allow cross-vendor and cross-domain comparisons in future work. We showed that complex MPPCA SVS performs better than NORDIC on these GE data in terms of contrast-to-noise gain and in the ability to resolve fine anatomical details post-denoising. This is different to previous results on comparing these two approaches with Siemens datasets [ 2 ]. These differences in trends are likely linked to the differences in MPPCA denoising methods used, with the MPPCA implementation from MRTrix (v3.0.7) [ 34 ] not including singular value shrinkage correction to compensate for noise-induced inflation of eigenvalues [ 21 ] or the option of patch averaging, as found in the DESIGNER implementation [ 35 , 36 ] that we used in our study. Additionally, cross-vendor differences in partial Fourier reconstruction could also explain some of these differences in denoising outcomes [ 50 ]. Siemens reconstruction often uses zero-filled or POCS, whereas GE most commonly uses Homodyne reconstruction. The downstream denoising consequences of vendor-specific partial Fourier reconstruction methods are not well explored. However, even within a single vendor, these choices have been shown to significantly affect denoising outcomes [ 50 ]. We compared the performance of proprietary ARDL denoising and reconstruction, which benefits from within scanner integration and is modality agnostic, to open-source reconstruction and dMRI denoising using the offline reconstruction pipeline and commonly used PCA-based denoising methods [ 10 , 11 , 13 ]. These results suggest that PCA-based denoising methods outperform ARDL in terms of identifying fine image details, contrast to noise ratio gains, and fiber orientation estimation. This is not surprising, given that ARDL is 2D. Interestingly, ARDL significantly outperforms other methods when comparing the maximally suppressed CSF signal intensity in the ventricles, showing significantly lower noise-floor bias with a distribution closer to zero, but not a corresponding improvement in the signal dynamic range or ADC profile in regions of high diffusion anisotropy. This suggests that, post-denoising, the signal in ventricles, which is often used as a proxy for the noise distribution [ 2 ], does not capture the biophysical complexity to yield a generalizable metric of bias reduction across tissues. However, ARDL also operates in the complex domain and the overall benefits in reducing noise-induced biases agree with benefits offered by PCA-based denoising approaches when applied to complex data. ARDL requires a 256×256 input matrix size which was achieved by embedding the acquisition matrix (160×160) into a zero-filled oversized matrix, followed by a post-ARDL spline interpolation to the acquisition matrix size for comparison against the open-source denoising methods. To ensure our trends were not driven by our custom post-ARDL interpolation, we carried out a comparison where all denoising methods were zero-filled to a 256×256 matrix size, using the same processing as the ARDL preprocessing, followed by the spline interpolation to the acquisition matrix size (160×160) prior to denoising. The trends in our results remained unchanged, as shown in Supplementary Figure S1. Additionally, the comparisons carried out in this paper only highlight the performance of ARDL on the ‘high’ setting, supplementary Figure S2 shows that this was the best performing mode available from scanner selection (low, medium, high). However, a limitation of this work is that we did not test ARDL capabilities beyond the ‘high’ setting, as ARDL uses a tuning hyperparameter ranging from 0 to 1, and the high setting corresponds to 0.75. Additionally, ARDL includes unringing features, which were not evaluated against open-source alternatives[ 51 ] in this study. Previous work has suggested that denoising earlier in the reconstruction process is beneficial to denoising outcomes [ 1 , 50 ], from the reduction of non-stationary g-factor contributions to noise [ 11 , 52 ]. Although ARDL benefits from denoising single-channel complex data rather than channel-combined data, it is important to note that 2D ARDL does not utilize the full redundancy of 4D dMRI data, unlike the open-source PCA-based approaches, developed explicitly for use in dMRI data. The proposed pipeline provides access to single-channel complex data, enabling evaluations of denoising earlier in the reconstruction process and across partial Fourier methods. 5. Conclusion Our offline reconstruction pipeline for GE dMRI acquisition was validated against scanner magnitude reconstruction and gives access to unfiltered complex domain data for acquisitions with in-plane and SMS acceleration. We found significant gains in dMRI data quality when using our offline reconstruction pipeline, allowing denoising in the complex domain, both for reducing noise-induced variance and bias. Open-source PCA-based (4D) methods outperformed proprietary 2D complex-domain denoising using deep learning (ARDL) in terms of qualitative resolution, reduction of noise-floor bias and variance. Acknowledgements FD is supported by a studentship funded by GE HealthCare and the University of Nottingham. SW and SNS are supported by an ERC Consolidator Grant (101000969). We would like to also thank Tim Sprenger, Jaemin Shin, Gavin Houston and Suchandrima Banerjee from GE HealthCare for useful discussions. Additionally, we would like to thank Martin Graves and Ilse Patterson for giving us access to the GE Premier scanner at CUH, while the Nottingham one was out of action. Funder Information Declared University of Nottingham, https://ror.org/01ee9ar58 GE Healthcare (United States), https://ror.org/01bwa4v12 European Research Council, https://ror.org/0472cxd90 , 101000969 References 1. ↵ Huynh K , Chang WT , Wu Y , Yap PT . Optimal shrinkage denoising breaks the noise floor in high-resolution diffusion MRI . Patterns . 2024 ; 5 ( 4 ): 100954 . doi: 10.1016/j.patter.2024.100954 OpenUrl CrossRef PubMed 2. ↵ Manzano Patron JP , Moeller S , Andersson JLR , Ugurbil K , Yacoub E , Sotiropoulos SN . Denoising diffusion MRI: Considerations and implications for analysis . Imaging Neurosci . 2024 ; 2 : 1 – 29 . doi: 10.1162/IMAG_A_00060 OpenUrl CrossRef 3. ↵ París G , Pieciak T , Jones DK , Aja-Fernández S , Tristán-Vega A , Veraart J. Thermal noise lowers the accuracy of rotationally invariant harmonics of diffusion MRI data and their robustness to experimental variations . Magn Reson Med . Published online 2025. doi: 10.1002/mrm.70035 OpenUrl CrossRef 4. ↵ Lemberskiy G , Baete SH , Veraart J , Shepherd T , Fieremans E , Novikov, Dmitry S. Achieving sub-mm clinical diffusion MRI resolution by removing noise during reconstruction using random matrix theory . In: 2019 ISMRM & ISMRT Annual Meeting . 2019:0770. Accessed November 3, 2025 . https://archive.ismrm.org/2019/0770.html 5. ↵ Ades-Aron B , Coelho S , Lemberskiy G , et al. Denoising Improves Cross-Scanner and Cross-Protocol Test–Retest Reproducibility of Diffusion Tensor and Kurtosis Imaging . Hum Brain Mapp . 2025 ; 46 ( 4 ): e70142 . doi: 10.1002/hbm.70142 OpenUrl CrossRef PubMed 6. ↵ Gudbjartsson H , Patz S. The Rician distribution of noisy MRI data . Magn Reson Med . 1995 ; 34 ( 6 ): 910 – 914 . doi: 10.1002/mrm.1910340618 OpenUrl CrossRef PubMed Web of Science 7. ↵ Jones DK , Basser PJ . “Squashing peanuts and smashing pumpkins”: How noise distorts diffusion-weighted MR data . Magn Reson Med . 2004 ; 52 ( 5 ): 979 – 993 . doi: 10.1002/mrm.20283 OpenUrl CrossRef PubMed 8. Hutchinson EB , Avram AV , Irfanoglu MO , et al. Analysis of the effects of noise, DWI sampling, and value of assumed parameters in diffusion MRI models . Magn Reson Med . 2017 ; 78 ( 5 ): 1767 – 1780 . doi: 10.1002/mrm.26575 OpenUrl CrossRef PubMed 9. ↵ Yao J , Tendler BC , Zhou Z , et al. Both noise-floor and tissue compartment difference in diffusivity contribute to FA dependence on b-value in diffusion MRI . Hum Brain Mapp . 2022 ; 44 ( 4 ): 1371 – 1388 . doi: 10.1002/hbm.26121 OpenUrl CrossRef PubMed 10. ↵ Cordero-Grande L , Christiaens D , Hutter J , Price AN , Hajnal JV . Complex diffusion-weighted image estimation via matrix recovery under general noise models . NeuroImage . 2019 ; 200 : 391 – 404 . doi: 10.1016/J.NEUROIMAGE.2019.06.039 OpenUrl CrossRef PubMed 11. ↵ Moeller S , Pisharady PK , Ramanna S , et al. NOise reduction with DIstribution Corrected (NORDIC) PCA in dMRI with complex-valued parameter-free locally low-rank processing . NeuroImage . 2021 ; 226 : 117539 . doi: 10.1016/J.NEUROIMAGE.2020.117539 OpenUrl CrossRef PubMed 12. ↵ Veraart J , Fieremans E , Novikov DS . Diffusion MRI noise mapping using random matrix theory . Magn Reson Med . 2016 ; 76 ( 5 ): 1582 – 1593 . doi: 10.1002/mrm.26059 OpenUrl CrossRef PubMed 13. ↵ Veraart J , Novikov DS , Christiaens D , Ades-aron B , Sijbers J , Fieremans E. Denoising of diffusion MRI using random matrix theory . NeuroImage . 2016 ; 142 : 394 – 406 . doi: 10.1016/J.NEUROIMAGE.2016.08.016 OpenUrl CrossRef PubMed 14. ↵ Manjón JV , Coupé P , Concha L , Buades A , Collins DL , Robles M. Diffusion Weighted Image Denoising Using Overcomplete Local PCA . PLOS ONE . 2013 ; 8 ( 9 ): e73021 . doi: 10.1371/journal.pone.0073021 OpenUrl CrossRef PubMed 15. Manjón JV , Coupé P , Buades A. MRI noise estimation and denoising using non-local PCA . Med Image Anal . 2015 ; 22 ( 1 ): 35 – 47 . doi: 10.1016/j.media.2015.01.004 OpenUrl CrossRef PubMed 16. ↵ Olesen JL , Ianus A , Ø stergaard L , Shemesh N , Jespersen SN . Tensor denoising of multidimensional MRI data . Magn Reson Med . 2023 ; 89 ( 3 ): 1160 – 1172 . doi: 10.1002/mrm.29478 OpenUrl CrossRef PubMed 17. ↵ Henriques RN , Ianuş A , Novello L , Jovicich J , Jespersen SN , Shemesh N. Efficient PCA denoising of spatially correlated redundant MRI data . Imaging Neurosci . 2023 ; 1 : 1 – 26 . doi: 10.1162/imag_a_00049 OpenUrl CrossRef 18. ↵ Marčenko VA , Pastur LA . Distribution of Eigenvalues for Some Sets of Random Matrices . Math USSR-Sb . 1967 ; 1 ( 4 ): 457 – 483 . doi: 10.1070/SM1967v001n04ABEH001994 OpenUrl CrossRef 19. ↵ Coupé P , Manjón JV , Gedamu E , Arnold D , Robles M , Collins DL . Robust Rician noise estimation for MR images . Med Image Anal . 2010 ; 14 ( 4 ): 483 – 493 . doi: 10.1016/j.media.2010.03.001 OpenUrl CrossRef PubMed 20. ↵ Does MD , Olesen JL , Harkins KD , et al. Evaluation of principal component analysis image denoising on multi-exponential MRI relaxometry . Magn Reson Med . 2019 ; 81 ( 6 ): 3503 – 3514 . doi: 10.1002/mrm.27658 OpenUrl CrossRef PubMed 21. ↵ Gavish M , Donoho DL . Optimal Shrinkage of Singular Values . IEEE Trans Inf Theory . 2017 ; 63 ( 4 ): 2137 – 2152 . doi: 10.1109/TIT.2017.2653801 OpenUrl CrossRef 22. ↵ Powell E , Schneider T , Battiston M , et al. SENSE EPI reconstruction with 2D phase error correction and channel-wise noise removal . Magn Reson Med . 2022 ; 88 ( 5 ): 2157 – 2166 . doi: 10.1002/mrm.29349 OpenUrl CrossRef PubMed 23. ↵ Lebel RM . Performance characterization of a novel deep learning-based MR image reconstruction pipeline . Published online August 2020. http://arxiv.org/abs/2008.06559 24. Park SI , Yim Y , Lee JB , Ahn HS . Deep learning reconstruction of diffusion-weighted brain MRI for evaluation of patients with acute neurologic symptoms . Sci Rep . 2024 ; 14 : 24761 . doi: 10.1038/s41598-024-75011-1 OpenUrl CrossRef PubMed 25. Nicolas Behl . Deep Resolve: Unrivaled Speed in MRI [White Paper] . April 2024 . Accessed November 4, 2025 . https://www.siemens-healthineers.com/en-uk/magnetic-resonance-imaging/technologies-and-innovations/deep-resolve 26. Peeters H , Chung H , Valvano G , et al. Philips SmartSpeed No compromise [White Paper] . July 2022 . Accessed November 4, 2025 . https://www.philips.co.uk/healthcare/technology/smartspeed-ai 27. ↵ GE Healthcare . AIR Recon DL: Simply better image quality . 2021 . Accessed November 4, 2025 . https://www.gehealthcare.com/products/magnetic-resonance-imaging/air-recon-dl 28. ↵ Haacke EM , Lindskogj ED , Lin W. A fast, iterative, partial-fourier technique capable of local phase recovery . J Magn Reson 1969 . 1991 ; 92 ( 1 ): 126 – 145 . doi: 10.1016/0022-2364(91)90253-P OpenUrl CrossRef 29. ↵ Noll DC , Nishimura DG , Macovski A. Homodyne detection in magnetic resonance imaging . IEEE Trans Med Imaging . 1991 ; 10 ( 2 ): 154 – 163 . doi: 10.1109/42.79473 OpenUrl CrossRef PubMed 30. ↵ Griswold MA , Jakob PM , Heidemann RM , et al. Generalized autocalibrating partially parallel acquisitions (GRAPPA) . Magn Reson Med . 2002 ; 47 ( 6 ): 1202 – 1210 . doi: 10.1002/MRM.10171 OpenUrl CrossRef PubMed Web of Science 31. ↵ Larkman DJ , Hajnal JV , Herlihy AH , Coutts GA , Young IR , Sta Ehnholm G. Use of Multicoil Arrays for Separation of Signal from Multiple Slices Simultaneously Excited . Published online 2001. doi: 10.1002/1522-2586(200102)13:2 OpenUrl CrossRef 32. Setsompop K , Gagoski BA , Polimeni JR , Witzel T , Wedeen VJ , Wald LL . Blipped-controlled aliasing in parallel imaging for simultaneous multislice echo planar imaging with reduced g-factor penalty . Magn Reson Med . 2012 ; 67 ( 5 ): 1210 – 1224 . doi: 10.1002/mrm.23097 OpenUrl CrossRef PubMed 33. ↵ Weaver JB . Simultaneous multislice acquisition of MR images . Magn Reson Med . 1988 ; 8 ( 3 ): 275 – 284 . doi: 10.1002/MRM.1910080305 OpenUrl CrossRef PubMed 34. ↵ Tournier JD , Smith R , Raffelt D , et al. MRtrix3: A fast, flexible and open software framework for medical image processing and visualisation . NeuroImage . 2019 ; 202 : 116137 . doi: 10.1016/j.neuroimage.2019.116137 OpenUrl CrossRef PubMed 35. ↵ Ades-Aron B , Veraart J , Kochunov P , et al. Evaluation of the accuracy and precision of the diffusion parameter EStImation with Gibbs and NoisE removal pipeline . NeuroImage . 2018 ; 183 : 532 – 543 . doi: 10.1016/j.neuroimage.2018.07.066 OpenUrl CrossRef PubMed 36. ↵ Chen J , Ades-Aron B , Lee HH , et al. Optimization and validation of the DESIGNER preprocessing pipeline for clinical diffusion MRI in white matter aging . Imaging Neurosci . 2024 ; 2 : imag-2-00125 . doi: 10.1162/imag_a_00125 OpenUrl CrossRef 37. ↵ Jenkinson M , Bannister P , Brady M , Smith S. Improved Optimization for the Robust and Accurate Linear Registration and Motion Correction of Brain Images . Published online 2002. doi: 10.1006/nimg.2002.1132 OpenUrl CrossRef PubMed 38. ↵ Jenkinson M , Smith S. A global optimisation method for robust affine registration of brain images . Med Image Anal . 2001 ; 5 ( 2 ): 143 – 156 . doi: 10.1016/S1361-8415(01)00036-6 OpenUrl CrossRef PubMed Web of Science 39. ↵ Sotiropoulos SN , Jbabdi S , Xu J , et al. Advances in diffusion MRI acquisition and processing in the Human Connectome Project . NeuroImage . 2013 ; 80 : 125 – 143 . doi: 10.1016/J.NEUROIMAGE.2013.05.057 OpenUrl CrossRef PubMed 40. ↵ Uğurbil K , Xu J , Auerbach EJ , et al. Pushing spatial and temporal resolution for functional and diffusion MRI in the Human Connectome Project . NeuroImage . 2013 ; 80 : 80 – 104 . doi: 10.1016/j.neuroimage.2013.05.012 OpenUrl CrossRef PubMed Web of Science 41. ↵ Andersson JLR , Sotiropoulos SN . An integrated approach to correction for off-resonance effects and subject movement in diffusion MR imaging . NeuroImage . 2016 ; 125 : 1063 – 1078 . doi: 10.1016/J.NEUROIMAGE.2015.10.019 OpenUrl CrossRef PubMed 42. ↵ Mohammadi-Nejad AR , Pszczolkowski S , Auer D , Sotiropoulos S. Multi-modal neuroimaging pipelines for data preprocessing . Published online June 2020. doi: 10.5281/zenodo.3909526 OpenUrl CrossRef 43. ↵ Bastiani M , Cottaar M , Fitzgibbon SP , et al. Automated quality control for within and between studies diffusion MRI data using a non-parametric framework for movement and distortion correction . NeuroImage . 2019 ; 184 : 801 – 812 . doi: 10.1016/J.NEUROIMAGE.2018.09.073 OpenUrl CrossRef PubMed 44. ↵ Sotiropoulos SN , Moeller S , Jbabdi S , et al. Effects of image reconstruction on fiber orientation mapping from multichannel diffusion MRI: Reducing the noise floor using SENSE . Magn Reson Med . 2013 ; 70 ( 6 ): 1682 – 1689 . doi: 10.1002/MRM.24623 OpenUrl CrossRef PubMed 45. ↵ Behrens TEJ , Woolrich M , Jenkinson M , et al. Characterization and propagation of uncertainty in diffusion-weighted MR imaging . Magn Reson Med . 2003 ; 50 ( 5 ): 1077 – 1088 . doi: 10.1002/mrm.10609 OpenUrl CrossRef PubMed Web of Science 46. ↵ Hernández M , Guerrero GD , Cecilia JM , et al. Accelerating Fibre Orientation Estimation from Diffusion Weighted Magnetic Resonance Imaging Using GPUs . PLOS ONE . 2013 ; 8 ( 4 ): e61892 . doi: 10.1371/journal.pone.0061892 OpenUrl CrossRef PubMed 47. ↵ Warrington S , Bryant KL , Khrapitchev AA , et al. XTRACT-Standardised protocols for automated tractography in the human and macaque brain . Neuroimage . 2020 ; 217 : 116923 . doi: 10.1016/j.neuroimage.2020.116923 OpenUrl CrossRef PubMed 48. ↵ Behrens TEJ , Berg HJ , Jbabdi S , Rushworth MFS , Woolrich MW . Probabilistic diffusion tractography with multiple fibre orientations: What can we gain? neuroimage . 2007 ; 34 ( 1 ): 144 – 155 . doi: 10.1016/j.neuroimage.2006.09.018 OpenUrl CrossRef PubMed Web of Science 49. ↵ Warrington S , Ntata A , Mougin O , et al. A resource for development and comparison of multimodal brain 3 T MRI harmonisation approaches . Imaging Neurosci . 2023 ; 1 : imag-1-00042 . doi: 10.1162/imag_a_00042 OpenUrl CrossRef 50. ↵ D’Antonio F , Warrington S , Morgan P , Sotiropoulos S. Single-Channel and Channel-Combined Complex Denoising for Diffusion MRI . In: 2025 ISMRM & ISMRT Annual Meeting . 2025 : 0694 . doi: 10.58530/2025/0694 OpenUrl CrossRef 51. ↵ Lee HH , Novikov DS , Fieremans E. Removal of partial Fourier-induced Gibbs (RPG) ringing artifacts in MRI . Magn Reson Med . 2021 ; 86 ( 5 ). doi: 10.1002/mrm.28830 OpenUrl CrossRef PubMed 52. ↵ Robson PM , Grant AK , Madhuranthakam AJ , Lattanzi R , Sodickson DK , McKenzie CA . Comprehensive quantification of signal-to-noise ratio and g-factor for image-based and k-space-based parallel imaging reconstructions . Magn Reson Med . 2008 ; 60 ( 4 ): 895 – 907 . doi: 10.1002/mrm.21728 OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted November 07, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Offline Reconstruction of Diffusion MRI Acquisitions for Comparison Between Complex PCA-based and AI-based Denoising Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Offline Reconstruction of Diffusion MRI Acquisitions for Comparison Between Complex PCA-based and AI-based Denoising Francesco D’Antonio , Shaun Warrington , Jose-Pedro Manzano-Patron , Paul S. Morgan , Stamatios N. Sotiropoulos bioRxiv 2025.11.07.682283; doi: https://doi.org/10.1101/2025.11.07.682283 Share This Article: Copy Citation Tools Offline Reconstruction of Diffusion MRI Acquisitions for Comparison Between Complex PCA-based and AI-based Denoising Francesco D’Antonio , Shaun Warrington , Jose-Pedro Manzano-Patron , Paul S. Morgan , Stamatios N. Sotiropoulos bioRxiv 2025.11.07.682283; doi: https://doi.org/10.1101/2025.11.07.682283 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Neuroscience Subject Areas All Articles Animal Behavior and Cognition (7618) Biochemistry (17636) Bioengineering (13861) Bioinformatics (41848) Biophysics (21401) Cancer Biology (18538) Cell Biology (25424) Clinical Trials (138) Developmental Biology (13354) Ecology (19861) Epidemiology (2067) Evolutionary Biology (24287) Genetics (15585) Genomics (22463) Immunology (17701) Microbiology (40300) Molecular Biology (17141) Neuroscience (88437) Paleontology (666) Pathology (2825) Pharmacology and Toxicology (4813) Physiology (7633) Plant Biology (15107) Scientific Communication and Education (2042) Synthetic Biology (4285) Systems Biology (9808) Zoology (2268)