Corneal nerve fiber size adds utility to the diagnosis and assessment of therapeutic response in patients with small fiber neuropathy

Small fiber neuropathy (SFN) is a common feature of many inflammatory diseases, often presenting with pain and disability. SFN is diagnosed using symptoms, thermal threshold testing, and intra-epidermal nerve fiber quantification. Corneal confocal microscopy (CCM) is an ophthalmic imaging technique which non-invasively quantifies corneal nerve fiber (CNF) density, branch density and length, and has comparable diagnostic and superior ability to identify nerve regeneration compared to skin biopsy. CNF size (width and area) depends upon the number of fibers within each nerve, as well as pathology (e.g., swelling), and may provide additional sensitivity to diagnose SFN and identify nerve repair. We have compared the utility of the standard CCM variables employed to CNF size in patients with diabetic sensorimotor polyneuropathy or sarcoidosis-associated SFN, and in patients with SFN following cibinetide administration, an agent which promotes nerve repair. The results show that: 1) CNF width distribution and area depend upon neuropathy severity; 2) CNF area, density, branch density and length possess comparable discriminatory power for diagnosing neuropathy; 3) CNF area is related to length by a quadratic function which is predictive for both healthy subjects and those with SFN; 4) CNF area is a useful variable for quantifying change in CNF morphology.

Comparison of CCM measurements obtained from control subjects versus those having diabetes or sarcoidosis The demographics of the groups studied are summarized in Table 1 . Although the data spread was relatively wide for each CCM variable, the median values obtained using both automated methods generally exhibited a progressive reduction in value with increasing Neuropathy Disability Score (NDS; Figure 2 ; Table 2 ). However, subjects without neuropathy (NDS 0–2) deviated from what is otherwise a negative monotonic curve, especially for NBD, such that this group was similar to the NDS 3–5 group. Adjustment for the significantly younger mean age (a weak determinant of corneal nerve fiber abundance 29 ) of the NDS 0–2 group using a linear model with age as a covariate did not affect this deviation (data not shown). The measurements obtained from the sarcoidosis group were also very similar to the NDS 3–5 group for most CCM variables, except for NFA WxL (Fig. 2D ), which surprisingly was similar to the distribution of the normal subjects. When expressed as percent of the mean value of control subjects (Fig. 2F ), no significant differences were noted between the neuropathy groups, except for NFA WxL in the sarcoidosis group which was significantly different from NFD (p = 0.005), NBD (p = 0.026), and NFA FIJI (p < 0.001). Further, the relative ranks of the means of each CCM variable expressed as percent of the control group mean were generally maintained across the groups, except for NFD and NFA WxL in the diabetic patients with NDS 0–2 and sarcoidosis subjects, respectively. Table 1 Demographics of subject populations. group n males age Disease duration (years) BMI HbA1c (%) NDS Diabetes a NDS (0–2) 21 15 37.1 (16.5) b,c 17.9 (15.1) 26.4 (5.5) 7.9 (1.3) 0.381 (0.740) Diabetes a NDS (3–5) 21 12 55.9 (11.0) 33.5 (16.7) 28.9 (3.8) 8.6 (1.4) 3.810 (0.928) Diabetes a NDS (6–8) 19 11 59.0 (11.3) 44.3 (11.0) 28.4 (3.8) 8.0 (1.1) 6.895 (0.875) Diabetes a NDS (9–10) 20 12 57.0 (14.6) 28.1 (15.1) 24.9 (7.2) 7.8 (2.1) 9.800 (0.410) Control 48 25 46.2 (16.9) NA 26.4 (5.5) 5.7 (0.3) 0.425 (1.231) Sarcoidosis 63 31 50.2 (10.0) 11.1 (9.3) 29.3 (5.5) ND ND Type 2 Diabetes d 48 30 63.1 (7.0) ND 30.7 (4.4) 7.2 (1.4) ND a Diabetes group 1. b Significantly younger than the other groups. c Mean (standard deviation). d Diabetes group 2. NA: not applicable; ND: not determined. Figure 2 Descriptive statistics of corneal confocal microscopy (CCM) variables ( A–E ) (median with interquartile range and maximum/minimum). In general, values of all variables decrease with increasing severity of neuropathy. The sarcoidosis group was generally similar to the NDS (0–2) and (3–5) groups for NFD, NBD, NFL, NFA FIJI, but diverged for NFA WxL. ( F ) Percent reduction from the corresponding means of normal subjects for each CCM variable (error bars: SEM) illustrates that small nerve fiber losses range from 80% to 40% with increasing severity of neuropathy. [NFD = nerve fiber density; NBD = main nerve branch density; NFL = nerve fiber length; NFA WxL = nerve fiber area determined by length times average width; NFA FIJI = nerve fiber area determined by the the total number of pixels within the nerve plexus]. Table 2 Corneal nerve fiber analysis variables. group NFD number/mm 2 NBD number/mm 2 NFL mm/mm 2 NFA WxL µm/mm 2 NFA FIJI µm/mm 2 Diabetes NDS (0–2) 21.5 (7.3) a 6.3–36.5 (22.5) b 21.6 (11.7) 8.8–50.0 (16.7) 12.8 (3.4) 6.1–19.0 (12.2) 39385 (10626) 19297–59771 (37461) 20966 (7718) 9792–33607 (20176) Diabetes NDS (3–5) 20.5 (9.5) 3.1–40. (18.8) 28.2 (17.3) 1.0–64.1 (22.9) 12.7 (5.0) 2.4–22.0 (11.7) 39843 (14474) 11383–67150 (36511) 22122 (10740) 4333–46624 (19694) Diabetes NDS (6–8) 15.2 (8.7) 1.0–28.8 (14.6) 17.2 (13.2) 0–51.8 (18.8) 9.6 (4.3) 2.4–15.3 (9.4) 30340 (12571) 7274–48140 (29335) 16414 (7594) 5234–29111 (16196) Diabetes NDS (9–10) 10.3 (8.5) 0–30.2 (8.2) 10.8 (10.7) 0–34.4 (7.5) 7.3 (4.7) 1.4–19.0 (6.2) 28827 (13603) 7217–58661 (20523) 12516 (7773) 3563–32794 (10573) Control 28.8 (4.8) 21.2–37.5 (28.1) 36.5 (14.4) 19.8–68.8 (32.3) 17.1 (3.2) 11.2–24.0 (16.4) 52331 (9344) 34573–73445 (50965) 30931 (8735) 19576–50238 (29600) Sarcoidosis 19.2 (5.4) 7.4–37.5 (19.4) 25.5 (11.8) 5.8–52.6 (25.0) 12.7 (2.9) 6.0–19.9 (12.5) 46890 (10737) 22247–74559 (45975) 19243 (5407) 8657–32188 (18921) a Mean (standard deviation). b Range (median). The NDS 0–2 and sarcoidosis groups were also characterized by having the lowest interclass correlation between CCM variables, irrespective of the variable assessed, compared to diabetic patients with mild and severe neuropathy (Fig. 3 ). Additionally, the interclass correlation coefficients between the different CCM variables were calculated for the controls, diabetes, and sarcoidosis study groups as summarized in Fig. 4 . NFD, NFL, NBD, NFA WxL, and NFA FIJI exhibited similar high correlations with each other, with Pearson correlation coefficients ranging from 0.83–0.91 for the control group, 0.84–0.96 for

The subject groups were further analyzed to determine the relationship between NFL and NFA FIJI, two variables calculated using the number of pixels contained within the identified image targets. A plot of NFL versus NFA FIJI pixel values obtained by analyzing the same images from 192 individuals (Fig. 7A ) shows that data obtained from the diabetic subjects with severe neuropathy is distributed within the left (lower) portion of the curve, while the sarcoid data set is restricted to the central portion of the curve, and control subjects within the right (upper) portion. Clearly the relationship is nonlinear as the slope of the curve becomes progressively steeper as corneal nerve fiber abundance increases. Linear regression analysis performed using the Passing-Bablok technique showed that the highly significant relationship between these two variables is given by the quadratic equation NFA FIJI = 2.78 × 10 −4 (NFL) 2 + 0.72(NFL) + 471 (solid curve in Fig. 7 ; p < 0.0001). This function predicts that differences between two nearby points on the region of the curve corresponding to more abundant nerve fibers will be larger for NFA FIJI than for NFL. For example, if a subset of subjects is established from the data set illustrated in Fig. 7A by arbitrarily selecting subjects with NFL pixels ≥ 1500 (i.e., ≥ 9.8 mm/mm 2 ), groups of 50 subjects with diabetes and 55 with sarcoidosis result, having an overlapping distribution. These groups are characterized by having a mean NFL (and SEM) of 13.79 (0.38) mm/mm 2 for diabetes and 13.35 (0.37) for sarcoidosis (a difference of ~3.2% of the mean value for the entire group), which is not statistically different (p = 0.409). In contrast, the corresponding mean NFA FIJI values are 23,238 (904) µm 2 /mm 2 for diabetes compared to 20,460 (862) µm 2 /mm 2 for sarcoidosis (a difference of ~12.7% of the mean value for the entire group), which is significantly different (p = 0.028). Alternatively, logistic regression performed using the same data set shows that these groups cannot be differentiated using NFL, in contrast to NFA FIJI which does provide significant discrimination. Finally, NFL and NFA FIJI data obtained from an independent diabetes group (diabetes group 2; n = 48) also fit this regression curve well (r 2 = 0.80; Fig. 7B ), providing cross-validation confirmation for this relationship. Figure 7 ( A ) The relationship between NFL and NFA FIJI is characterized by a quadratic function. The change in slope of this curve shows that NFA changes more rapidly than NFL as nerve fiber densities approach normal (right portion of the curve). Conversely, differences bewteen these two variables are more similar in cases of severe neuropathy (left portion of the curve). ( B ) CCM data obtained using an independent (cross-validation) group of Type 2 diabetic subjects fall on the prediction curve from panel A.

These data confirm that CCM using automated image analysis is an excellent methodology for quantification of corneal nerve morphology and provides a reliable estimate of small nerve fiber abundance. All CCM variables evaluated demonstrated equivalent ability to discriminate patients with diabetic neuropathy from healthy controls, based on comparable AUCs of the receiver operating curves, a finding which is similar to the results of a previous study using ACCmetrics which observed AUCs within the ~0.75 range 15 . Additionally, the cut off points used to distinguish neuropathy patients with diabetes from normal controls are similar to those determined by Petropoulos et al . 7 for subjects with diabetic neuropathy established using nerve conduction velocity, Neuropathy Disability Score, and vibratory-thermal sensory thresholds to diagnose the presence of neuropathy. Moreover, these results demonstrate that additional useful information is contained within each CCM image in the form of the width distribution of nerve fiber bundles and total nerve fiber area, which also vary as a function of neuropathic severity. Therefore, in addition to a reduction in nerve fiber length, density and branch density, the relative loss of nerve fiber bundles in the 2.8–4.0 µm range appears to be an additional metric to characterize small fiber neuropathy. Thus, with increasing corneal nerve fiber loss, a preferential drop out of nerve fiber bundles around the mean width was observed, shifting the width frequency distribution towards thicker fiber bundles. It is interesting that the relative proportion of the very small fibers (<2.7 µm) is not reduced as much as fibers in the midrange of width as shown by the non-symmetrical curves around the normal mean (Fig. 6 ). It is possible that the fiber bundles of smaller widths which appear to be relatively preserved represent a thinning process arising from the time-dependent drop out of individual nerve fibers from wider bundles. Alternatively, or in addition to nerve fiber drop out, the enrichment of wider fiber bundles could occur via individual fiber swelling, which is a known characteristic of small fiber neuropathy 21 . Additional longitudinal study and electron microscopy level histology correlated with the CCM findings will be required to investigate these possibilities further. The data also suggest that the differences in the caliber of nerve fiber bundles may allow CCM methodology to identify specific types of neuropathy. In this regard, the comparison between diabetes- and sarcoidosis-associated SFN is interesting. In the present study, NFD, NBD, and NFL of the sarcoidosis subjects overlapped with the NDS (0–2) and NDS (3–5) diabetes groups. However, although both groups had a similar proportional reduction in nerve fibers around the normal mean (i.e., in the 2.7–4.0 µm range), a large relative increase in wider nerve fiber bundles was noted in the sarcoidosis group (Figs 6 and S2 ). This characteristic of the sarcoidosis group contributed to the poorer interclass correlation coefficients noted when compared to the diabetes group. Further studies are required to establish whether both thin and thick corneal nerve fiber bundles drop out to a more balanced extent in subjects with mild diabetic neuropathy compared to subjects having sarcoidosis, or whether sarcoid neuropathy is characterized by a preponderance of smaller nerve fiber bundle loss. Additionally, it will be interesting to determine whether nerve fiber bundle width distributions in SFN of other etiologies are different or similar to the two neuropathy groups studied here. Notably, the relationship between NFL and NFA FIJI in control subjects or patients with neuropathy fall along the same regression line, indicating that the morphological characteristics of the corneal nerve fiber plexus are predictably related to nerve fiber abundance, regardless of the presence or absence of neuropathy or disease severity. Although the individual CCM variables were highly intercorrelated, the values corresponding to subjects with milder neuropathy were lower than those for subjects with severe neuropathy. This finding is consistent with a variable impact of the disease process on different morphological components of the nerve fiber plexus, but also may reflect the relatively crude assessment of neuropathic deficits using NDS, especially for milder neuropathy. Therefore, although each CCM variable is equally useful determine the existence of severe neuropathy, NFL and NFA appear to be more sensitive discriminators of the presence of mild neuropathy. In this respect, however, it is important to note that NFL is a one-dimensional variable (length), whereas NFA is two-dimensional (length times width). Therefore, NFA provides additional information, especially for increasing complexity of corneal nerve plexus, i.e., for states of mild neuropathy. These data are supportive of the conclusion that NFA FIJI may have particular utility f

Corneal confocal images of the sub-basal nerve layer of the central cornea were acquired using the Rostock Corneal Module of the Heidelberg Retinal Tomograph III (HRT III) using a previously established protocol 34 . For the diagnostic group (diabetes versus healthy control subjects), 5–8 images (median 6) of corneal nerves from the sub-basal layer of one eye were selected on the basis of technical quality by an expert (MF) blinded as to subject identity. 579 images were selected and analyzed and the results averaged for each individual. For the longitudinal assessment group (sarcoidosis patients), as the SFNL associated with this disease has been reported to be “patchy” in distribution 35 , both eyes were imaged and 2479 images were selected based on image quality by an expert (MB) blind to treatment group (6–18; median of 10 per eye corresponding to baseline and the 28 day follow up period). Data obtained from the images from each eye were averaged and the mean of both eyes used for analyses. The image data obtained by the HRT III are derived using an intentionally very limited axial resolution (depth of focus) of 7.6 µm 36 . The sub-basal nerve plexus which contains the nerve bundles consists of a single, very thin (<10 µm) layer parallel to the corneal surface 37 . Therefore, each technically acceptable image likely samples the entire sub-basal layer. To reduce the unlikely possibility that images could be acquired at slightly different corneal depths within the sub-basal plexus, a standard protocol for acquiring images was used, and the results of multiple independent images averaged. Automated data analysis was performed using two different algorithms: ACCmetrics and a custom-developed macro for FIJI. ACCmetrics 15 provides individual estimates of corneal nerve fiber density, corneal nerve branch density and corneal nerve fiber length. Nerve fiber width distribution is also generated as the frequency of occurrence of specific fiber widths (ranging from 1 to a maximum of 8 pixels) determined perpendicularly to each pixel constituting the long axis of identified nerve fibers 15 . NFA was calculated as the total number of width pixels multiplied by the total number of length pixels (NFA WxL). Although the ACCmetrics algorithm improves target identification, the image processing steps utilized degrade and de-emphasize fine structural details such as nerve fiber beading and subtle changes in nerve fiber width, and additionally bridge small gaps along the length of each nerve fiber. Therefore, a second algorithm was developed to determine NFA. This method employs a custom-developed macro for FIJI (NFA FIJI), a public-domain image analysis program, version 1.47e. The program creates a binary image of the corneal nerve fibers by applying a Gaussian filter to enhance target contrast and adjusts detection threshold based on background brightness, followed by use of the eigenvalues of the Hessian matrix to identify tubular nerve fiber structures 38 . NFA FIJI is defined as the image area (µm 2 ) covered by the resulting mask, and is expressed as per mm 2 of corneal surface area.

The datasets generated and analyzed in the current study are available from the corresponding author on reasonable request.