What has brain diffusion MRI taught us about chronic pain: a narrative review

Information source

PubMed and Scopus databases were interrogated for articles dated up to 10^th March 2022. Articles were also searched using the Medical Subject Headings (MeSH) term on PubMed.

Review strategy

This review follows the recommendations from the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR) Checklist. Search strategies were developed with a librarian of the Health Sciences Library of the Université de Sherbrooke. The keywords chosen for the review were: “magnetic resonance imaging”, “diffusion”, “pain”, and “brain”. The full search strategy can be viewed in the supplementary material (supp 1.).

Study strategy and inclusion/exclusion criteria

The process for selecting studies in this investigation is depicted in Figure 1. First, from a collection of 447 papers, we discarded reviews and any articles that did not explicitly mention “diffusion,” “brain,” or “pain” in their title or abstract. Second, we eliminated case studies, non-English articles, and those in which pain was not the main focus. Third, we removed studies if they did not employ diffusion MRI (dMRI) to examine brain regions in individuals suffering from chronic pain or if the study population included participants under the age of one, veterans, amputees, recipients of deep brain stimulation (DBS), individuals with traumatic brain injury (TBI), postoperative patients, stroke or neurodegenerative disease-induced chronic pain sufferers, or subjects experiencing experimentally induced acute pain. Methodological articles and animal studies were also excluded. Fourth, we separated the remaining articles into three categories: chronic primary pain, chronic secondary pain based on the IASP classification ^48,49, or articles missing information about the chronic pain condition, and only included articles on chronic primary pain. We further excluded studies that limited dMRI analysis solely to peripheral nerves, those that included only healthy control subjects, or those that failed to report results (a note on acute pain was also added in supplementary material [supp 2.]). Finally, an additional article was incorporated post hoc; this article had been referenced multiple times in the selected literature and was identified as pertinent due to its focus on the use of dMRI in chronic pain, despite not being retrieved by our initial search terms in its title or abstract.

• Download figure
• Open in new tab

Figure 1.

Flowchart of the article selection process, 58 final articles were included for the analyzes in this review.

Study selection and analyses

The first three steps described in the section “study strategy and inclusion and exclusion criterias” were conducted by MS and PT; any disagreement was resolved between MS and PT. Subsequently, the classification into primary chronic pain and secondary chronic pain was made by MS and supported by PT and GLe. The remaining articles were then separated into five categories based on the IASP classification: (i) chronic widespread pain; (ii) complex regional pain syndrome; (iii) chronic primary headache or orofacial pain; (iv) chronic primary visceral pain and (v) chronic primary musculoskeletal pain. If applicable, half of the articles in each category were separated and respectively analyzed by MS and PT. The remaining articles were analyzed by MM and GLe. Finally, each reviewer (MS, PT, MM and GLe) extracted, based on the chart developed by PT, general study design information and dMRI-specific study data.

Chronic primary headache or orofacial pain

The IASP classification defines chronic primary headache or orofacial pain as headache or orofacial pain manifesting for at least 15 days per month and persisting for over three months. The duration of untreated daily pain is a minimum of two hours or may manifest in multiple shorter episodes ⁴⁸. In reviewing this category, 21 articles were identified: 11 focused on migraine ^58–68, two on cluster headache ^69,70, one on chronic headache ⁷¹, three on temporomandibular disorder (TMD) ^72–74 and four on burning mouth syndrome ^75–78.

In the studies examining migraine, five investigations employed FSL-TBSS for voxel-wise statistical analyses on whole-brain fractional anisotropy (FA) skeletons ^58,64–67. While Neeb et al. ⁶⁴ and Coppola et al. ⁶⁵ solely relied on TBSS, Kattem Husøy et al. ⁵⁸ and Szabó et al. ⁶⁷ incorporated additional tractography analyses. Gomez-Beldarrain et al. ⁶⁶ further extended their approach to include a region-of-interest (ROI) analysis, utilizing both white and gray matter atlases. Neeb et al. reported no significant differences in FA, mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD) between migraine patients and healthy controls. Coppola et al. examined both episodic and chronic migraines, finding no FA differences in episodic migraine patients but identifying higher RD and MD in bilateral superior and posterior corona radiata (CR), the bilateral genu of the corpus callosum (CC), the bilateral posterior limb of the internal capsule (IC) bilateral superior longitudinal fasciculus (SLF) in chronic migraine patients. Kattem Husøy et al. discovered elevated AD in major tracts, most notably in the bilateral CC, corticospinal tract (CST), inferior fronto-occipital fasciculus (IFOF), inferior longitudinal fasciculus (ILF), and left SLF. They also noted lower volumes in the CC and IFOF when employing deterministic tractography, but only in the new-onset headache group. Szabó et al. found lower FA and higher MD and RD values in frontal white matter bundles. Probabilistic tractography, originating from TBSS seed regions, revealed these bundles to be connected to the orbitofrontal cortex, insula, thalamus, and dorsal midbrain. Gomez-Beldarrain et al. observed lower FA in the TBSS skeleton and identified significantly different TBSS clusters between controls and migraine patients, which were mapped to ROIs in the Johns Hopkins University (JHU) diffusion tensor imaging (DTI)-based Atlas and the MNI structural Atlas. The interior insula, bilateral cingulate gyri, and right uncinate fasciculus were pinpointed as regions with lower FA values.

Three investigations exclusively utilized tractography ^60,63,68. Chong et al. ⁶⁰ employed probabilistic tractography to segment and average white matter microstructural properties with the aim of differentiating migraine patients from those with post-traumatic headaches. Their predictive model achieved an accuracy rate of 78%; however, no diffusion metrics were compared. Planchuelo-Gomez et al. ⁶³ used probabilistic tractography to build a whole brain connectome by measuring streamline counts between gray matter regions. They observed both higher and lower numbers of streamlines in connections involving specific regions such as the superior frontal gyrus. Additionally, they found variations in FA, AD, and RD in connections involving regions like the hippocampus. Silvestro et al. ⁶⁸ took a similar approach, employing probabilistic tractography to construct a whole-brain connectome. They quantified the connection probability between gray matter regions and further examined the resultant connectome using network-based statistics (NBS) and graph theory network analysis. Their findings highlighted nodes with significantly higher connection probabilities in multiple regions, including the precuneus, cuneus, amygdala, calcarine cortex, and posterior cingulate cortex, anterior cingulate cortex (ACC), postcentral gyrus, lingual and fusiform gyri, middle frontal gyrus and inferior and superior parietal lobules.

Three investigations utilized voxel-based analysis (VBA) ^59,61,62. Zhang et al. ⁵⁹ reported no significant differences using a whole-brain VBA approach via SPM12. On the other hand, Marciszewski et al. ⁶¹ observed elevated MD in various brain regions, including the spinal trigeminal nucleus, dorsal medial/lateral pons, midbrain periaqueductal gray (PAG), and cuneiform nucleus. They also identified higher FA in the white matter regions of the medial lemniscus and ventral trigeminal thalamic tract. Dasilva et al. ⁶² further extended their approach to include a ROI analysis, utilizing both white and gray matter atlases. They discovered lower FA values in specific patient subgroups: the ventro-lateral PAG was affected in migraine patients without aura, while the ventral trigemino-thalamic tract showed reduced FA in migraine patients with aura.

In the studies examining cluster headache, two investigations employed FSL-TBSS for voxel-wise statistical analyses on whole-brain FA skeletons ^69,70. Szabo et al. ⁶⁹ reported a significant increase in MD, AD, and RD across widespread white matter regions in the frontal, parietal, temporal, and occipital lobes. They also found reduced FA in the CC and in specific frontal and parietal white matter tracts, such as the bilateral forceps minor and major, the bilateral CR, the IC and external capsule (EC), the cerebral peduncle, the parietal juxtacortical white matter and the IFOF, predominantly on the contralateral side of the pain. Notably, AD exhibited a negative correlation with the frequency of headache attacks. Teepker et al. ⁷⁰ reported alterations in FA in multiple brain regions, including the brainstem, thalamus, IC, superior and inferior temporal regions, frontal lobe, occipital lobe, and cerebellum.

In one study examining chronic headache, conducted by Miller et al. ⁷¹, they used deterministic tractography to segment and average diffusion metrics on tracts. They observed higher FA in the cingulum.

In the studies examining temporomandibular disorders (TMD), two investigations employed FSL-TBSS for voxel-wise statistical analyses on whole-brain FA skeletons ^72,73. Both studies extended their analysis to include tractography. Moayedi et al. ⁷³ went further by also incorporating VBA. Salomons et al. ⁷² conducted their analysis using TBSS and then segmented major tracts via probabilistic tractography. They observed that FA values in connected white matter tracts along the CST were associated with feelings of helplessness; however, no group comparison of diffusion metrics were reported. Moayedi et al., after performing TBSS analyses, employed probabilistic tractography, using seed regions identified from significant TBSS clusters, to measure connection probability. They also applied a whole-brain VBA. Their findings indicated lower FA in the anterior limb of the IC and the EC. Additionally, they observed a higher connection probability from the CC to the frontal pole and a lower connection probability from the CC to the dorsolateral prefrontal cortex.

One investigation exclusively utilized whole-brain VBA ⁷⁴. Gustin et al. ⁷⁴ observed no significant differences in FA within the primary somatosensory cortex (S1), nor did they find any correlation between FA and reorganization of S1.

In the studies examining burning mouth syndrome (BMS), two investigations employed FSL-TBSS for voxel-wise statistical analyses on whole-brain FA skeletons ^76,77. Khan et al. ⁷⁶ extended their study by incorporating probabilistic tractography to segment major white matter tracts. They found no significant differences in either TBSS-based diffusion metrics or average tract metrics and volumes. Tan et al. ⁷⁷, who solely employed TBSS, also found no significant differences in the whole-brain FA skeleton.

Two investigations exclusively utilized tractography to construct whole brain connectomes ^75,78. Both used probabilistic tractography to measure the streamline count between gray matter regions and subsequently applied graph theory network analysis methods to further examine the connectomes. Wada et al. ⁷⁵ observed localized changes in connectivity within the ACC and prefrontal cortex, specifically in the medial orbitofrontal cortex and pars orbitalis. They also noted strengthened connections between the ACC and medial prefrontal cortex with regions such as the basal ganglia, thalamus, and brainstem. Despite these findings, no significant differences were observed in graph theory metrics. Kurokawa et al. ⁷⁸ reported increased connectivity and betweenness centrality in the left insula, right amygdala, and right lateral orbitofrontal cortex as well as reduced betweenness centrality and connectivity in the right inferotemporal cortex.

Chronic primary visceral pain

The IASP classification defines chronic primary visceral pain as persistent or recurrent pain (for longer than three months) that occurs in the internal organs of the head (or neck) region and of the thoracic, abdominal, or pelvic cavities, unexplained by any other condition ⁴⁸. The pain anatomical location corresponds with typical referral pain patterns from specific internal organs. It can manifest in diverse forms, such as pain in the digestive system, the thoracic region, and the abdominal region, as well as pelvic pain originating from the viscera of the digestive, urinary, and genital systems. In this category, 19 studies were included. Among them, four focused on primary dysmenorrhea (PDM) ^79–82, seven on irritable bowel syndrome (IBS) ^83–89, four on urologic chronic pelvic pain (UCPP) ^90–93, two on prostatitis/chronic pelvic pain syndrome (CCP/CPPS) ^94,95, one on cystitis/bladder pain syndrome ⁹⁶ and one on provoked vestibulodynia (PVD) ⁹⁷.

In the studies examining primary dysmenorrhea (PDM), two investigations employed FSL-TBSS for voxel-wise statistical analyses on whole-brain fractional anisotropy (FA) skeletons ^79,82. While both studies utilized TBSS, Liu et al. ⁸² extended their analysis to include tractography. Dun et al. ⁷⁹ observed lower FA values, alongside higher mean diffusivity (MD) and radial diffusivity (RD), across various white matter fibers tracts. These tracts included the splenium part of the corpus callosum (CC), the posterior limb of the internal capsule (IC), the superior and posterior of the corona radiata (CR), as well as the posterior thalamic radiation. Conversely, Liu et al. reported higher FA and lower MD and RD values, primarily in the CC, fornix, bilateral IC, bilateral external capsule (EC), CR, and bilateral posterior thalamic radiation, bilateral sagittal stratum, right cingulum and bilateral superior longitudinal fasciculus (SLF). In addition, tractography was employed by Liu et al. to visually inspect connectivity originating from TBSS seed regions.

Two investigations exclusively utilized tractography ^80,81. Both studies performed tractography in a population atlas to manually segment white matter tracts of interest. Subsequently, these segmentations were registered to the native space for a region-of-interest (ROI) analysis of diffusion tensor imaging (DTI) metrics. He et al. ⁸⁰ reported lower FA values in connections between the thalamus and somatosensory cortex (S1) as well as between the thalamus and insula. Conversely, they found higher FA values in the tracts connecting the thalamus to the dorsal anterior cingulate cortex (ACC) and the supplementary motor area. Liu et al. ⁸¹ observed lower FA and axial diffusivity (AD), along with higher RD and MD, in a specific cluster located in the dorsal posterior cingulum and in the parahippocampal segment of the cingulum bundle.

In the studies examining irritable bowel syndrome (IBS), three investigations employed FSL-TBSS for voxel-wise statistical analyses on whole-brain FA skeletons ^86–88. Chen et al. ⁸⁷ further refined their TBSS analysis with a ROI approach using multiple atlases, including the Johns Hopkins University (JHU) white matter atlas, Harvard-Oxford cortical structural atlas, MNI structural atlas, Talairach Daemon labels, and the Juelich histological atlas. Similarly, both Hubbard et al. ⁸⁶ and Nan et al. ⁸⁸ employed an ROI-based approach, albeit solely using the JHU white matter atlas for segmentation. Chen et al. observed higher FA values in the fornix and EC adjacent to the right posterior insula. Hubbard et al., while not finding any significant differences in the whole FA skeleton, did report lower FA in the lower dorsal cingulum; however, they found no variations in MD and RD values. Nan et al. reported lower FA and higher RD specifically in the genu of the CC, with no observed differences in MD.

Two studies exclusively employed tractography methods ^83,89. Irimia et al. ⁸³ utilized deterministic tractography to calculate average tract DTI metrics while Liu et al. ⁸⁹ applied tractography originating from the bilateral posterior cingulate gyrus to compute metrics such as average tract FA, fiber length, and streamline count. Irimia et al. observed higher FA values in white matter bundles innervating the S1. In contrast, Liu et al. did not report any significant differences in their measured parameters.

Two studies employed voxel-based analysis (VBA) ^84,85. Ellingson et al. ⁸⁴ extended their analysis by incorporating a ROI approach using the JHU atlas and applied probabilistic tractography to compute the number of streamlines between certain atlas ROIs. Qi et al. ⁸⁵ similarly extended their analysis by applying tractography but originated it from fMRI-defined ROI clusters to compute average tract DTI metrics, streamline count, and path length. Ellingson et al. observed higher FA values in various grey matter regions including the globus pallidus, putamen, medial thalamus, and sensory and motor cortices as well as in white matter regions such as the primary cortical projections from the thalamus, posterior cingulate, frontal lobe and ACC white matter and CC when using VBA. They also reported higher MD within the globus pallidus, IC, thalamus, CR, and areas connected to sensory, pre-frontal, and posterior parietal regions. In terms of tract density, they found higher values in tracts connecting the thalamus to the prefrontal cortical regions and the medial dorsal nuclei to the ACC. Conversely, they observed lower tract density in connections between the globus pallidus and the thalamus. In contrast, Qi et al. did not find any significant differences in path length, tract count, or FA within the fibers connecting the bilateral ventral ACC to the inferior parietal lobules.

In the studies examining urologic chronic pelvic pain (UCPP), one study conducted by Huang et al. ⁹¹ utilized FSL-TBSS for voxel-wise statistical analyses on whole-brain FA skeletons. Huang et al. extended their TBSS analysis by incorporating a ROI approach using the JHU white matter atlas. Huang et al. reported lower FA and higher AD values in the thalamic radiation but found no significant differences in either MD or RD.

Three studies employed VBA ^90,92,93. The 2018 study by Woodworth ⁹² focused solely on whole-brain VBA, while the 2015 Woodworth study ⁹⁰ extended its analysis to include probabilistic tractography. Alger et al. ⁹³ incorporated a ROI approach using the JHU white matter atlas. In the 2018 study, Woodworth reported significant correlations between DTI measures and urinary protein quantifications; however, no comparative analyses were conducted between different groups in terms of diffusion metrics. The 2015 Woodworth study found lower FA, lower generalized FA, lower tract density, and higher MD in brain regions typically associated with the perception and integration of pain information. Interestingly, the study did not compare IBS patients to a control group but did find them to be significantly different from UCPP patients. Alger et al. primarily aimed to evaluate the variability in FA measurements. Their study obtained data from various acquisition sites for neurologic chronic pain syndrome in healthy controls but did not conduct any comparative analyses on diffusion metrics.

In the studies examining prostatitis/chronic pelvic pain syndrome (CCP/CPPS), one study utilized FSL-TBSS for voxel-wise statistical analyses on whole-brain FA skeletons⁹⁴. Farmer et al. ⁹⁴ reported no significant differences in whole-brain FA skeletons across various DTI measures.

In a single study conducted by Huang et al. ⁹⁵, both VBA and tractography were employed to calculate whole-brain connectome and graph theory metrics. Huang et al. observed lower global efficiency in the right middle frontal gyrus (orbital part) and higher global efficiency in the left middle cingulate and paracingulate gyrus. Additionally, they reported increased local efficiency in the left middle cingulate and paracingulate gyri, as well as the paracentral lobule.

In the study examining cystisis/bladder pain syndrome, a study utilized FSL-TBSS for voxel-wise statistical analyses on whole-brain FA skeletons ⁹⁶. Farmer et al. ⁹⁶ reported lower FA values in specific regions: the right thalamic radiation, the left forceps major, and the right longitudinal fasciculus. Conversely, they observed higher FA in the right SLF as well as in the bilateral inferior longitudinal fasciculus (ILF).

In the study examining provoked vestibulodynia (PVD), a study conducted by Gupta et al. employed VBA on the whole brain and a ROI approach with the Harvard-Oxford subcortical atlas to segment gray matter regions ⁹⁷. The investigators reported extensive increases in FA within the somatosensory and basal ganglia regions as well as variations in MD specifically in the basal ganglia.

Chronic primary musculoskeletal pain

The IASP classification defines chronic primary musculoskeletal pain as persistent or recurrent (for longer than three months) pain that occurs in the muscles, bones, joints, or tendons ⁴⁸. A classic example is chronic primary low-back pain. This category is further subclassified based on pain anatomical location, including the upper back as chronic primary cervical pain, the mid-back as chronic primary thoracic pain, the lower back as chronic primary low-back pain, and the limbs as chronic primary limb pain. In this category, ten studies were included. Among them, five focused on sub-acute/chronic low back pain (SBP/CLBP) ^34,98–101, three on chronic musculoskeletal pain syndrome ^102–104, one on nonspecific low back pain ¹⁰⁵ and one on chronic neck pain ¹⁰⁶.

In the studies examining sub-acute/chronic low back pain (SBP/CLBP), four studies employed FSL-TBSS to conduct voxel-wise statistical analyses on whole-brain fractional anisotropy (FA) skeletons ^98–101. Among these, Kim et al. ¹⁰¹ solely employed TBSS, Mansour et al. ⁹⁸ extended their analysis to include tractography, and both Ma et al. ⁹⁹ and Ceko et al. ¹⁰⁰ applied a region of interest (ROI) approach. Specifically, Ma et al. utilized a white matter atlas for their ROI analysis, while Ceko et al. employed ROI clusters derived from a prior fMRI study. Kim et al. observed that CLBP patients exhibited reduced FA in both the somatosensory cortex (S1)-back and S1-finger regions when compared to controls by averaging FA skeleton diffusion tensor imaging (DTI) metrics in ROIs. Mansour et al. found lower FA values in three distinct clusters: one in the temporal part of the superior longitudinal fasciculus (SLF), a second in the left retro-lenticular part of the internal capsule (IC), and a third involving the left anterior limb of the IC as well as portions of the corpus callosum (CC), including the anterior corona radiata (CR) in SBP patients with pain that persisted compared to SBP patients that recovered after a year. Ma et al. reported lower FA in several regions: the CC, bilateral anterior and right posterior thalamic radiation, right SLF, and left anterior CR among CLBP patients. Contrarily, Ceko et al. didn’t find any whole-brain DTI metric differences in the FA skeleton, but instead measured an increased FA in the left insula post-treatment.

One study exclusively used probabilistic tractography to reconstruct the whole brain connectome and evaluate the number of connections and connection probabilities ³⁴. Vachon-Presseau et al. ³⁴ observed higher density of connections between corticolimbic regions, such as the nucleus accumbens, the amygdala, the hippocampus and the prefrontal cortex in patients with pain that persisted compared to SBP patients that recovered after a year. No comparisons were made with respect to DTI metrics.

In the studies examining non-specific chronic low back pain (NSCLBP), Pijnenburg et al. ¹⁰⁵ study exclusively used probabilistic tractography to reconstruct the whole brain connectome and evaluate graph theory network analysis. The investigators observed lower local efficiency in NSCLBP cases; notably, no comparative analysis was conducted on diffusion metrics.

In the studies examining chronic musculoskeletal pain, two studies employed FSL-TBSS to conduct voxel-wise statistical analyses on whole-brain FA skeletons ^102,103. Lieberman et al. ¹⁰³ extended their analysis by incorporating a ROI approach using a white matter atlas. Bishop et al. ¹⁰² also employed an ROI approach using the JHU white matter atlas, a whole brain fixel-based analysis and a probabilistic tractography with constrained spherical deconvolution (CSD) reconstruction of the whole-brain connectome to apply network-based statistics (NBS). Lieberman et al. found no significant differences in FA or axial diffusivity (AD) in the whole-brain TBSS FA skeleton. However, they reported higher radial diffusivity (RD) in multiple regions, including the body of the CC, right SLF and both anterior and posterior limbs of the IC. Further ROI analyses revealed lower FA in the splenium of the CC and the left temporal lobe branch of the cingulum bundle adjacent to the hippocampus. Elevated RD was also found in several regions, such as the splenium of the CC, the right limbs of the IC, part of the external capsule (EC) adjacent to the insular cortex, the SLF and the cerebral peduncle. Bishop et al. observed lower F1 values in the right EC, right SLF, and right uncinate fasciculus, as well as lower F2 values in the left cingulum and the splenium of the CC (F1 and F2 are respectively the first and second fiber population partial volume fractions, these diffusion properties can only be calculated when using advanced diffusion models that can account for multiple crossing fibers within a voxel). They also found lower mode of anisotropy in several areas including the splenium of the CC, left cerebral peduncle, and bilateral EC. In their whole-brain TBSS skeleton analysis, no differences in F1 and F2 were found. However, fixel-based analysis revealed reduced fiber density in the splenium of the CC and the right temporal lobe white matter region of the inferior frontal-occipital fasciculus (IFOF). No differences were observed in fiber cross-section. Utilizing a connectome approach, Bishop et al. reported increased connectivity between several regions, including the hippocampus, parietal cortex, thalamus, precuneus, and visual cortex structures like the calcarine and cuneus gray matter. One study exclusively used an ROI approach, utilizing DTI and neurite orientation dispersion and density imaging (NODDI) metrics ¹⁰⁴. By targeting ROIs within both white matter and gray matter atlases, Cruz-Almeida et al. ¹⁰⁴ observed lower orientation dispersion index values in the white matter of several regions, including the anterior CR, right posterior thalamic radiation, uncinate fasciculus, superior cerebellar penduncle, and fornix.

In a study focused on chronic neck pain, conducted by Coppieters et al. ¹⁰⁶, an ROI approach using a white matter atlas was employed to assess DTI metrics. The investigation revealed no significant differences in ROI-based DTI metrics between patients with chronic neck pain and healthy controls.

Chronic widespread pain

The IASP classification defines chronic widespread pain as persistent or recurrent (for longer than three months) diffuse musculoskeletal pain that occurs in a minimum of four body regions and in at least three of four body quadrants (upper–lower/left–right side of the body) ^48,107. In this category, six studies were included. All of which focused on fibromyalgia (FM) ^108–113.

In the studies focused on fibromyalgia, one study conducted by Ceko et al. ¹⁰⁹ utilized FSL-TBSS for voxel-wise statistical analyses on whole-brain fractional anisotropy (FA) skeletons. Additionally, they employed tractography to examine connectivity originating from significant TBSS clusters. While no differences were observed in whole-brain FA skeletons with respect to diffusion tensor imaging (DTI) measures, the investigators did report lower FA in regions adjacent to areas where significant gray matter volume differences were identified.

One study exclusively used tractography to construct whole brain connectomes using streamline counts between gray matter regions ¹¹⁰. Kim et al. ¹¹⁰ used probabilistic tractography to build a whole brain connectome by measuring streamline count between gray matter regions. Notably, they found no differences in the white matter fiber count connecting areas associated with hyperalgesia and clinical pain (no local diffusion metrics were compared).

One study exclusively used a whole brain voxel-based analysis (VBA) ¹¹¹. Hadanny et al. ¹¹¹ observed higher FA in several regions, including the anterior thalamic radiation, left insula, right thalamus, and superior thalamic radiation.

Three studies exclusively used region of interest (ROI) based analysis methods ^108,112,113. Sundgren et al. ¹⁰⁸ used ROI masks for the whole brain and manually placed ROI spheres to utilize a histogram comparison method. Their findings included lower FA in the right thalamus, with no significant results for apparent diffusion coefficient (ADC). Fayed et al. ¹¹² also manually positioned ROI spheres to compare FA and ADC metrics but found no significant differences between groups. Lutz et al. ¹¹³ manually segmented ROIs to compare FA and ADC metrics. They observed lower FA in both the thalami, thalamocortical tracts, and both insular regions, alongside higher FA in the postcentral gyri, amygdala, hippocampi, superior frontal gyri, and anterior cingulate gyri. No significant differences were found in ADC metrics.

Complex regional pain syndrome (CRPS)

The IASP classification defines complex regional pain syndrome (CRPS) as persistent, or recurrent (for longer than three months) pain characterized by its regional distribution and time course. The pain typically begins distally in an extremity following trauma and is disproportionate in both magnitude and duration when compared to the usual course of pain after similar tissue injuries ^48,114,115. While two subtypes of CRPS have been identified, they are beyond the scope of this review. In this category, two studies were included ^116,117.

In the studies focused on Complex regional pain syndrome (CRPS), two studies employed FSL-TBSS for voxel-wise statistical analyses on whole-brain fractional anisotropy (FA) skeletons ^116,117. While both studies used TBSS, Geha et al. ¹¹⁷ extended their approach to include probabilistic tractography, which was applied to investigate connectivity originating from voxel-based morphometry (VBM) seed regions. Geha et al. reported a lower FA cluster within the left callosal fiber tract. Hotta et al. ¹¹⁶, who solely utilized TBSS, observed higher mean diffusivity (MD), axial diffusivity (AD), and radial diffusivity (RD) in the genu, body, and splenium of the corpus callosum (CC) as well as in the left anterior, posterior and right superior parts of the corona radiata (CR). Across the whole-brain FA skeletons, they observed average lower FA, higher MD and RD, and no differences were observed in AD.

Summary of our findings

To illustrate the results covered in this review, we have provided a summary of the types of chronic pain, the analysis methods and the reported regions and tracts used for the study of chronic pain with diffusion MRI in figure 3.

• Download figure
• Open in new tab

Figure 3.

An interactive visual summary of the reviewed studies in terms of the types of chronic pain studied (A), the used analysis methods (B) and the tracts and regions reported as significant findings (C). In blue: primary headache and orofacial, red: primary visceral pain, green: primary musculoskeletal pain, purple: chronic widespread pain and orange: complex regional pain syndrome. Dynamic figure can be found here: https://osf.io/4wyqt/

Acquisition parameters

Diffusion MRI metrics, though quantitative, are somewhat restricted in terms of their sensitivity and specificity. This necessitates careful interpretation of these metrics, considering methodological, technical, and biological factors. A notable issue affecting the consistency and reliability of dMRI is the substantial variation in scanning parameters across different studies. For instance, a recent study explored the variability in diffusion-weighted MRI across multiple sites, scanners, and subjects. The study demonstrated that, under specific acquisition conditions, the variability between different scanners assessing the same subject could be comparable to the variability between different subjects assessed on the same scanner ¹²⁰. As for any MRI sequences, many parameters need to be properly chosen to ensure high-quality diffusion-sensitive images. Three key parameters significantly influence both dMRI image quality and the outcomes of their subsequent analyses: 1) the b-value; representing the strength, duration and timing of the diffusion encoding gradients ^43,121; 2) the number of diffusions encoding gradient directions; generally representing the number of diffusion gradient directions applied over a sphere ^122,123; and 3) the voxel dimensions; representing the length, width and height of the 3D image voxels ¹²⁴. For reference, the most common acquisition parameters used for DTI are in the range of a b-value of 1000 s/mm², 30 unique gradient directions and a 2mm isotropic resolution.

1. Across studies, the b-value parameter was the most stable with ∼75% of studies using a b-value of 1000 s/mm². However, some studies reported b-values in an unusual range— such as 700, 800, 900, 1200, and 1300 s/mm²—without providing any explanation. Additionally, certain studies failed to report the b-value at all. Strikingly, only two groups performed multi-shell acquisitions, an acquisition strategy introduced over 15 years ago ¹²⁵, which allows for state-of-the-art dMRI analysis. Using modern multiband sequences, multi-shell images for whole brain in-vivo imaging can be acquired in an acceptable amount of time -- between 10 and 20 minutes at most.

2. When investigating the number of diffusion encoding gradient directions, we found that approximately 55% of the studies acquired a minimum of 30 directions (with a maximum of 99 directions). This parameter, however, exhibited significant variability; some studies used as few as six or nine directions. This is of concern because the theoretical minimum required to adequately describe a diffusion tensor is six gradient directions ^126,127. In the broader context of tractography and fiber orientation estimation, high angular resolution diffusion imaging (HARDI) has emerged as an effective acquisition strategy designed to address the limitations of traditional diffusion tensor imaging (DTI) ^128,129. Based on the acquisition of over 50 gradient directions at a single high b-value, HARDI (even when paired with a multi-shell acquisition) can be completed in a reasonable timeframe, between 10 and 20 minutes at most.

3. Lastly, when investigating image resolution, the overall voxel volume ranged from 2.4mm³ to 20mm³. While forty-seven percent of studies used the conventional 2mm isotropic resolution, there was notable variability, with some studies acquiring highly anisotropic voxels -- such as 1.875×1.873×3 or 2×2.5×4mm³. These anisotropic voxel dimensions can introduce biases into diffusion magnetic resonance imaging (dMRI) analyses. For instance, larger voxels are more likely to contain multiple fiber populations, thereby reducing fiber orientation homogeneity and affecting diffusion metrics ¹²⁴. Additionally, anisotropic voxels can influence tractography algorithms, particularly in situations involving branching fibers¹³⁰.

Overall, the observations made in this section highlight the heterogeneity of acquisition parameters and pre-processing pipelines across chronic pain dMRI studies. This diversity underscores a critical challenge for the field: the need for a more unified approach that would facilitate comparative and cumulative research. For a recent review of dMRI preprocessing, we refer to article ¹³¹and for a more detailed characterization of the impact of MRI acquisition parameters on diffusion models, we refer to international benchmark competitions reports, notably that discuss the impact of different inversion time (TI) and echo time (TE) ^132–134.

Processing and analysis tools

The tools provided by the FMRIB group at Oxford were the most used, with over 70% of the papers using one of the FSL tools to process their dMRI data. Almost 50% of the papers use the Tract Based Spatial Statistic (TBSS) pipeline ¹³⁵ to identify voxel-wise differences. Although TBSS is a valid approach that provides significant advantages over classic whole-brain voxel-based analysis when it comes to group comparison, this method does not use the full potential of dMRI images as it summarizes the complexity of the whole-brain white matter into a WM skeleton that is only a few voxels wide. Also, this approach does not completely exclude registration errors ¹³⁶. Therefore, almost no information coming from tracks spanning up to the cortex or tracks that are smaller or located in complex regions can be found in the TBSS maps. The other most commonly used software to process dMRI data were ExploreDTI ¹³⁷ and MRtrix ¹³⁸, both offering advanced tools specifically designed for dMRI data to generate tractography and diffusion metrics maps. To circumvent issues brought by TBSS, some studies presented a clever approach to identify white matter tracts impacted by the chronic pain condition under study. They used the clusters of significant differences identified in the TBSS results as seeds to perform probabilistic tractography. Although this approach allows for the identification of actual white matter tracts that were not present in the WM skeleton, it still cannot reconcile the fact that a significant volume of white matter was not included in the original TBSS analysis. Similarly, several groups identified GM ROIs from fMRI experiments and then expanded this region to include adjacent white matter as a WM ROI to extract diffusion properties. These approaches need to be interpreted carefully as a WM bundle passing close to a GM region does not necessarily connect with that region. Indeed, many WM tracts travel long distances in the brain without connecting with each region they are bordering along the way. A more appropriate approach could have been to use the GM ROI to generate seeds from/to which WM fibers might connect and extract anatomically plausible tracks from a whole brain tractogram.

To illustrate the constraints of the TBSS methods, we examined the disparities between TBSS and a track-based approach for two specific tracks, namely the fornix and the accumbofrontal (AcF) track, both of which are of particular significance in chronic pain research owing to the regions they connect. Notably, the fornix serves as the primary pathway for efferent signals from the hippocampus, and it plays a critical role in memory circuitry ¹¹⁹. Moreover, hippocampus volume has been shown to be a risk factor in the transition from acute to chronic pain ³⁴. The AcF track connects the orbitofrontal cortex to the nucleus accumbens ¹³⁹, both regions that are also implicated in the transition from acute to chronic pain ³⁰. To demonstrate the benefits of employing targeted approaches for investigating white matter tracks involved in chronic pain, we analyzed the overlap of two streamline bundles extracted using separate techniques with a commonly used whole brain FA skeleton (method described in supplementary material [supp 3.]). As a quantitative measure of overlap, voxels intersected by these tracks were extracted and the percentage of voxels overlapping the binarized FA skeleton were output for each track. The fornix and AcF tracks are displayed in Figure 7 along with the percentage of overlapping voxels showing 15% voxels overlap between TBSS and tractography approached for the fornix (figure 7 left), and only 12% voxels overlap for the AcF track (figure 7 right). This analysis was reproduced for a few other subjects from the OpenPain database and similar overlaps were obtained (not shown).

• Download figure
• Open in new tab

Figure 7.

A visualization of the overlap between the FSL FA skeleton used commonly in whole brain diffusion MRI analysis and two different reconstructed white matter bundles taken from two different tractography analysis approaches. The population average right fornix (Fx) bundle used in the RecobundleX analysis pipeline is displayed in blue with regions overlapping the FA skeleton displayed in yellow. Voxels overlapping the bundle were extracted and the percentage of these voxels overlapping the FA skeleton are displayed as a bar graph. The same metrics are displayed for the Left Accumbofronal (AcF) track which was generated in native imaging space by extracting tracks that traversed ROIs (pink) from the nucleus accumbens and the orbital frontal gyrus. Only a small proportion of these tracks overlap the whole brain FA skeleton from FSL, suggesting that more targeted analysis approaches may be more sensitive to detecting microstructural alterations in some white matter structures.

Where should the field go?

Similar to the challenges faced by fMRI for processing and analysis ^127,130, dMRI is facing reproducibility and replication issues. As evidenced by the results of this review, there is a substantial degree of variability in published findings, methodologies, and metrics across different studies. While there exists several good guidelines discussing future directions for neuroimaging-based pain biomarker research ^3,25,140,141, we suggest specific strategies to address dMRI’s challenges: i) updating dMRI signal modelling, ii) increase access and availability of dMRI data for CP, and iii) adopting standardized pipelines specific to dMRI.

• i) Updating dMRI local reconstruction methods: To identify and increase group differences, novel biomarkers that increase sensitivity and specificity must be found. Recent advancements in magnetic resonance (MR) hardware and acquisition schemes— such as high angular resolution diffusion imaging (HARDI), diffusion spectrum imaging (DSI), and multi-shell protocols—as well as analysis methods enable the exploration of such brain markers ¹⁴². However, the field of chronic pain (CP) has yet to yield the full potential of these methods. For instance, diffusion tensor imaging (DTI) has been so prevalent that it is often used interchangeably with dMRI in literature. More recent studies in dMRI have moved away from DTI, as it fails to accurately represent multiple fiber populations within a single white matter voxel. The tensor model typically falls short in capturing accurate microstructural information in voxels where fiber crossing occurs; given that this happens in approximately 60 to 90% of all white matter voxels in the brain ¹⁴³, the DTI method is increasingly seen as providing anatomically unsound information – especially when applying tractography. Recent dMRI research has begun to move away from DTI, adopting newer methods aiming for a more precise representation of underlying white matter tissue organization. Within the studies reviewed here, four studies employed whole-brain connectome metrics and two utilized multi-compartment local models. While these newer methods are promising, a caveat is warranted: if HARDI and multi-shell protocols gain widespread acceptance without standardization, there is a risk of exacerbating variability and noise in the CP literature. Notably, the application of these methods on DTI data will most likely provide unreliable results. Therefore, we advocate for the adoption of these advanced techniques only within a structured framework of standardized protocols, reproducible analytical pipelines and cross-validation.

• ii) Increasing access and availability of dMRI data for pain: One of the notable challenges in the study of chronic pain using diffusion magnetic resonance imaging (dMRI) is the limited access and availability of extensive open-access datasets. These limitations hamper the statistical power and generalizability of research findings. As such, studies included in this review had a maximum of 103 participants and a minimum of 7 participants (with an average of 37.33 ± 27.39 participants). Recent advancements in the democratization of machine learning methods, in data sharing ^144,145, in data harmonization ^146,147, and the adoption of uniform metrics present significant opportunities for addressing this issue. By leveraging these advancements, researchers have the potential to rapidly expand and enrich open-access datasets specifically focused on chronic pain.

• iii) Adopting standardized pipelines specific to dMRI: Both the field of chronic pain research and diffusion magnetic resonance imaging (dMRI) are subject to large intra-group variability. For chronic pain, such variability can be attenuated by refining the specificity of clinical evaluations, thereby reducing confounds related to the chronic pain condition itself. One approach could be selecting study participants with consistent clinical criteria or ensuring a large enough sample size for data-driven selection. For dMRI, variations can be minimized at every stage, from data collection to analysis. It is important to note that dMRI has specific challenges, unlike other types of structural MR imaging. These challenges include image susceptibility distortions, region-of-interest positioning, image registration, and data smoothing. Conventional structural MR analysis techniques, such as whole-brain voxel-based analysis, might be less suitable due to common dMRI pitfalls ¹⁴⁸. Therefore, we recommend that the acquisition parameters, pre-processing, processing and analysis be conducted with standardized pipelines specific to dMRI.

Overall, the goal of this review was not to favor one method over another but to provide an overview of the current state of the field. However, in writing this review, we emphasize the difficulty of finding commonalities amidst the diverse methods used for image acquisition, analysis, and communication. Consequently, we pinpoint specific areas that require attention and potential improvements. Ultimately, we hope that addressing these challenges will allow pain researchers to capture more reproducible, specific and subtle white matter abnormalities.