Preprocessing.

As fMRI signals are often corrupted by random noise, it is important to have these components appropriately modeled. fMRI data typically undergo preprocessing to remove physiological artifacts and validate model assumptions. The major steps involve motion correction, realignment, coregistration, normalization, and smoothing.

In this study, participants with a head movement exceeding 2 mm of maximum translation in any of the x, y, or z axes or 2^o of maximum rotation about the three axes were excluded from this study to minimize movement artifacts. The fMRI sequence was interleaved. A total of 76 volumes was scanned for each individual, and the first volume was discarded to allow for signal equilibrium of the initial MRI signals and adaptation of the participants to the circumstances. The remaining 75 consecutive volumes were used for data analysis. The fMRI data series were analyzed using Statistical Parametric Mapping 8 (SPM8) and Data Processing Assistant for Resting-State fMRI as described previously [13–15]. Briefly, the data series were aligned to the first image in each scan sequence and resampled by interpolation to correct voxel values, thus correcting motion artifacts. This procedure was repeated for each individual. Next, the functional data were coregistered to the T1-weighted anatomical image. Each participant’s anatomy was registered to a template brain provided by the Montreal Neurological Institute. Mapping was used to wrap the input image on to the template image, to produce a normalized image that could be compared between participants. The resampled voxel volume of the normalized images was 3-mm isotropic. Subsequently, smoothing was performed using a Gaussian kernel with a 4-mm full width at half maximum to increase the signal-to-noise ratio within the region. Conditional specific effects were estimated using the general linear model (GLM) in SPM8, which was constructed by convolving a boxcar sequence with the hemodynamic response function. A high-pass filter with a cutoff period of 128 s was used to remove low-frequency noise.

After the above preprocessing steps, the data series satisfied Gaussian distribution and were subjected to statistical analysis. In addition, slice timing correction was also performed [9,16].

Data analysis.

Interaction analysis was used to compare brain activities between the DPN, NDPN, and HV groups.

SPM8 was applied to compare the brain activation pattern in the three groups to establish its relationship with stimulation (cluster size k > 20; p < 0.05). One-way analysis of covariance with a GLM was used to analyze the interaction effect and bivariate covariance, with age and sex treated as covariant factors, followed by post-hoc analysis for differences between conditions. T0_task, T10_task, T34_task, and T44_task represented the fMRI data achieved under 0, 10, 34 (skin temperature), and 44°C stimulation. T0_off, T10_off, T34_off, T44_off represented the taking-off of these stimuli. A 3 × 3 bivariate covariance model was established with factor 1 set as DPN, NDPN, and HV, factor 2 set as (T0_task − T0_off) − (T34_task − T0_off), (T10_task − T10_off) − (T34_task − T0_off), (T44_task − T44_off) − (T34_task − T0_off) and change in skin temperature in response to the stimuli. All multiple comparisons were uncorrected, but set at a threshold of p < 0.05 and a cluster size of > 20 voxels.

The demographic features of the three groups were analyzed using different methods based on their distribution. Age, education, and body mass index (BMI) were analyzed using analysis of variance (ANOVA). The duration of DM, DPN, hypertension, and dyslipidemia, as well as the hours of exercise per day, were analyzed using a nonparametric test because of their non-normal distribution. The relationship between smoking history, alcohol consumption, and DM were tested using a likelihood ratio chi-squared test. MNSI score was compared between the DPN and NDPN groups using the t test. VAS score was evaluated by ANOVA. Statistical significance was set at p < 0.05. All analyses were performed using SPSS 17.0 software (SPSS Inc., Chicago, IL, USA).

Part 1. Activation pattern corresponding to pain sensation induced by thermal stimulus.

In the interaction analysis, DPN patients showed stronger activation than NDPN patients or HV in cerebral areas that participate in somatosensory pathways of pain sensation (i.e., the right insular cortex, left caudate nucleus, frontal gyrus, and cingulate cortex; Figs 3 & 4A). This finding is consistent with previous studies, which have demonstrated that pain perception is a matrix composed of many interacting functional areas in the brain [17]. Augmented cerebral activities in the primary (S1) and secondary (S2) somatosensory cortices, lateral frontal lobe, cerebellum, anterior cingulum, and thalamus have been observed in DPN patients [7,13]. These brain areas are considered responsible for distinguishing the position and intensity of painful stimuli [7]. Besides, decreased gray matter volume in cerebral regions related to somatic sensation and significant shrinkage of the spinal cord have been documented in DPN and subclinical DPN patients [18,19]. However, in this study, despite activation of these pain-related cerebral areas, DPN patients did not experience significantly more pain than HV and NDPN participants (Table 3). There could be several reasons for this. First, the thalamus is regarded as the gateway through which pain sensation enters the brain; it amplifies pain messages transmitted from peripheral neurons [20]. Hypervascularity and hyperexcitability of the thalamus has been demonstrated in painful DPN, whereas hypovascularity and hypoexcitablity are features of painless DPN [21]. In our study, no thalamus activation was observed in DPN patients, which may explain why no significant pain sensation was reported. Second, nerve damage in DPN is irreversible and progressive [22]. Decreased ability to constrict blood vessels [23] and greater impairment of the nerve axon reflex mediated by C fibers in response to arousal stimuli [24] are more likely to be seen in painless DPN than in painful DPN. As nerve damage progresses, pain increases until the nerve is so severely damaged that pain signals are no longer transmitted to the CNS. The result is that pain sensation decreases and the feet begin to feel numb. This is consistent with our results, which show decreased VAS results in DPN patients compared with in NDPN patients, although without statistical significance. Third, although cerebral areas recognized to be part of the pain matrix were activated in patients with painless DPN in our study, it may be that the pattern, sequence, and strength of cerebral activation was different from that evident in painful DPN. Previous studies have shown that painful DPN responds more strongly to acute thermal stimulus than painless DPN [25]. The activation in these brain areas may be insufficient to produce pain in patients with painless DPN.

In addition to the cerebral areas involved in the somatosensory pathway, more advanced cognition-related areas, including the left hippocampus and left fusiform gyrus, were also activated in DPN patients in response to 10°C stimulation (Fig 4B). The hippocampus is not only responsible for memory and learning [26], but is also related to spatial orientation [27], whereas the fusiform gyrus, which is located in the visual cortex, has the function of face recognition [28]. Our study suggests that advanced cognitive abilities such as learning, memory, spatial orientation, and face recognition may be affected in DPN patients. This result, which has not been reported before, is a major finding of this study. A previous meta-analysis demonstrated cognitive decline in NDPN patients [29], and it has been suggested that dysregulation of glycemic control is responsible for this cognitive impairment [30]. Our study confirms the involvement of the CNS in DPN, and suggests that the changes in these cerebral areas may precede clinically evident cognitive decline. Moreover, although we only stimulated the right legs of the participants, both the right and left cingulum were activated (Fig 4C), implying interconnections between cingulate nerve fibers. This explains why some patients have bilateral symptoms although their EMG findings only indicated unilateral dysfunction.

On comparing the activation patterns between NDPN patients and HV in response to 0°C stimulation, we found that activation in cerebellar vermis 1–3 was more prominent in NDPN patients (Fig 2B). The cerebellar vermis is composed of the anterior lobe vermis, posterior lobe vermis, and posterolateral region of the posterior lobe vermis. The main function of the vermis is to process proprioceptive and exteroceptive afferent signals and modulate muscle tone [31]. Thus, NDPN patients without complaints of peripheral neuropathy may still have subclinical changes in these areas.

This study has some limitations. Overall, activation in the right temporal lobe, right insula, left caudate nucleus, and frontal gyrus was significantly greater in DPN patients than in HV (Fig 3). However, when the response to each different temperature stimulus was examined separately, there was no significant difference between DPN patients and HV. There are several possible explanations for this finding. For example, the statistical analysis was influenced by the threshold cluster size of > 20. It is possible that some significant differences were masked when cluster size was limited to > 20. Moreover, it is possible that we missed the most sensitive temperature point in DPN patients. Stimulation with more distinctive temperatures should be performed in future studies.

Part 2. Activation pattern corresponding to the itching sensation induced by thermal stimulus.

Interaction analysis showed stronger activity in the cerebellar vermis and cingulum in NDPN and DPN patients than in HV (Figs 2 & 3). Both the cerebellar vermis and cingulum are involved in the recognition of an itching sensation [21,31]. Activation of the cerebellar vermis can lead to both pain and itching [31]. However, no patients in the DPN group reported itching in response to thermal stimulation. One reason for this could be that the temperature stimulus used in this experiment was insufficient to produce itching. Moreover, itching reflects a complicated network of brain activity. The hippocampus, amygdala, and subcortical cerebral areas are all involved in the itching network [32]. The activation of various other brain areas, such as the thalamus, anterior supplementary motor area, anterior insula, parietal inferior gyrus, and anterolateral prefrontal area, as well as the suppression of the orbitofrontal, medial prefrontal, medial cingulate, and primary motor cortex, are also related to the histamine-stimulated scratch response [33]. Therefore, activation of only the cerebellar vermis or cingulum may be insufficient to produce clinical itching. Because interactions exist between the peripheral and central pathways in the pathogenesis of DPN, it is necessary to explore whether peripheral neuropathy leads to activation of the above brain areas, or whether cortical activation results in peripheral neuropathic pain/itching. Although fMRI has many limitations and cannot always be performed in clinical settings because of its cost, it has been widely used to detect the neurological changes in pediatric epilepsy, Alzheimer’s disease, and other neurodegenerative diseases [34–36]. Our study demonstrates the utility of fMRI as a noninvasive and radiation-free tool for the diagnosis of DPN.