2.1.

Study Design and Patients

The evaluation reported in Sec. 3 includes data recorded during a prospective trial from a group of 14 healthy volunteers and 16 patients with RA. However, only the data of 12 healthy volunteers could be included in further analysis. Two volunteers have been excluded because their fingers were either accidentally contaminated with ICG or their finger nails had a fluorescent nail polish.

The imaging equipment was installed in a clinical environment and measurements on patients and volunteers were performed according to the approval of the ethics committee. An informed consent was obtained from each participant. All patients and volunteers underwent a clinical examination before an optical examination. To perform optical investigations, a dose of 0.1 mg/kg body weight ICG was administered intravenously, at a dose which is lower than that used for other diagnostic applications.^{16, 17}

As described by Mothes,¹⁷ fluorescence angiography using ICG has been proposed as a helpful indicator of free-flap perfusion in tissue transfers to the outer body. The dye is approved for use in cardiac and liver function diagnostics (Pulsion Medical Systems AG, Munich, Germany).

Table 1 summarizes the data of volunteers and patients with rheumatoid arthritis. It is worth noting that there is a strong difference in age between volunteers and patients. It is not clear to us whether this circumstance may have an impact or not.

Fig. 6

ROC curves computed from normalized variances of the correlation function shown in Fig. 4. The inset table contains particular values derived from the ROC curves.

Table 1

Comparison of gender and age between volunteers and patients.

4.1.

Normalization

The dye ICG is mostly bound to blood plasma proteins,¹⁰ and therefore, confined to the vascular system. The fluorescence of ICG recorded with our imaging system is therefore a measure of the vascularization in skin and tissue. However, the fluorescence intensity emitted from the surface of the hand is a two-dimensional (2D) projection of fluorescence photons from different depths in the hand. The dynamics of the ICG concentration at the point of observation is strongly dependent on the physiology of the blood circulation of the investigated person. Therefore, the measure is relative and not directly comparable between individuals. Already in the group of volunteers, fluorescence intensities of joints recorded as described in Sec. 3, show large variations in the time courses such as onset or maxima of fluorescence intensity. The onset time t ₀ (see Fig. 2) of the strong increase in fluorescence intensity after the bolus injection in the PIP-ROIs varied, e.g., between 30 and 55 s. All recorded data sets showed a detectable increase in fluorescence intensity, always later than 20 s after bolus injection. Therefore, the onset time t ₀ was defined as the time when the sliding average over 3 s (12 data points) of the intensity is increased by more than 30% over the average signal of the first 20 s. At bolus arrival, the intensity increases within 1 or 2 s by 1 order of magnitude. The time t _max between bolus injection and occurrence of the maximum of fluorescence intensity in the PIP joints ranges between 80 and 240 s. The corrected maxima of fluorescence intensity are between 1600 and 5700 counts.

These strong individual blood flow properties have to be corrected to allow the comparison of joints from volunteers and patients with RA. Therefore, a normalization of the fluorescence intensities measured in the ROIs is necessary. For this purpose, we used the finger tips, an area unaffected by rheumatoid arthritis, as a measure of the input function of fluorescence intensity close to the finger joints. Due to the small spatial separation between the finger tip and corresponding joints, only the local differences in the dye concentration are of importance. Other effects such as the exponential decay of the fluorescence intensity caused by the ICG liver clearance are negligible.

To further reduce the influence of the individual blood circulation, in particular to get rid of the individual time between bolus injection and its appearance in the ROIs, the time t ₀ was always used as a reference starting point for further evaluation of each time series, i.e., all data points with times t < t ₀ have been removed, and the time scale was shifted to t ^′ = t − t ₀. Furthermore, the length of the analyzed time series has been restricted to 480 s.

4.2.

Time Dependent Properties and Correlation

We defined a joint status based on fluorescence intensity ratios between the values of a joint and its corresponding finger tip. In a first approach, based on the experiences from our previous work,⁸ different relative fluorescence parameters of such types (i.e., intensities of ROIs normalized to that of the finger tip) have been calculated from the raw data of the time series. For instance, the maxima, the means, and the increasing and decreasing slopes, have been computed and evaluated.

In the present study, these approaches did not very well discriminate volunteers and patients, when compared with the individual clinical diagnosis. Figure 3 shows, as an example, the cumulative frequency of the normalized means of fluorescence intensities of PIP-joints of volunteers and patients. The curves for volunteers and patients strongly overlap. Subsequently, sensitivity and specificity as a function of the normalized mean have been derived from the cumulative frequency plots shown in Fig. 3, and plotted as receiver operating characteristics (ROC) curves. A ROC curve visualizes the characteristic interplay of sensitivity and specificity depending on a discriminator as a parameter. The diagram indicates a certain sensitivity for a given false positive rate (1-specificity). The coordinates of each point in the diagram are sensitivity and 1-specificity for the same discriminator. A value close to 0.5 for the area under the ROC-curve (ROC-AUC) means a complete overlap of the two groups to be distinguished. The closer the value of the AUC comes to 1, the better the two groups are separated. An AUC-value close to 1 means the graph has a sensitivity of 100% almost everywhere.

Fig. 3

Normalized cumulative frequencies versus normalized means of fluorescence intensities from PIP joints of healthy volunteers and RA patients.

The AUC of the ROC curve, based on sensitivity and specificity derived from the data of Fig. 3, resulted in AUC-values below 0.7. We concluded from these results that the ratio analysis taking into account simple time-integrated values, or values at certain times, is of poor discriminative power, since the available time evolution has been not considered. The more the dynamics of intensities in the finger tip and joints differ, the more likely it is that the joint is affected.

As an improvement, we suggest, therefore, to include an analysis of the time information by correlating the time courses of the fluorescence intensities of a joint and its corresponding finger tip.

The correlation function [TeX:] $\rho _{{\textit {\scriptsize FT}},{\textit {\scriptsize Joint}}} (\Delta t)$ ${\rho }_{\mathit{FT},\mathit{Joint}}\left(\Delta t\right)$ for a time shift Δt between the time-series of fluorescence intensities in the finger tip I _FT(t _n) and a joint [TeX:] $I_{Jo{\mathop{\rm int}} } (t_n)$ $I_{Jo\mathrm{int}}\left(t_n\right)$ is defined as:

As for the treatment of raw data ratios mean, maxima, and variance of the correlation function have been analyzed and yielded the best results for the variance. The bolus distribution and liver uptake effects have been eliminated as before by dividing the variances for the joints with one of the corresponding finger tips:

Table 2

Number of investigated joints in patients (32) including the clinical findings for swelling and/or pain at the time point of NIR imaging.

In Fig. 4, the cumulative frequency of normalized variances of the correlation function for all PIP joint groups of volunteers and patients are plotted. It can be seen that the normalized values for volunteers show lower values than those for patients.

Fig. 4

Normalized cumulative frequency versus normalized variances of the correlation function from PIP joints of healthy volunteers and RA patients.

It is obvious from Fig. 4, however, that the groups of volunteers and patients still overlap. However, the notched box plots presented in Fig. 5 show that the medians for volunteers and patients are clearly separated. These findings support that—within the evaluated group of individuals—our method based on variances of normalized correlation functions, allows to discriminate joints of RA patients from the joints of volunteers. For RA patients, a good separation between asymptomatic and inflamed joints was not possible. It must be noted that the latter result is mainly a consequence of the “gold standard” we had to compare with, but probably not a deficiency of our method. The comparison with inclusion criteria for RA (our quasigold standard) gave a result for all joints per patient only, i.e., all joints of RA patients were treated as having potentially inflamed joints without further differentiation. The worst separation with a poor AUC = 0.71 was, e.g., obtained for the group of PIP joints from the small fingers because a fairly high number of these joints (17 out of 32) was without acute swelling or pain, but classified as RA because of the overall assessment.

Fig. 5

Notch box plots of normalized variances of the correlation function from PIP joints of healthy volunteers (filled boxes) and RA patients (open boxes). Two medians are significantly different if their lower and upper confidence intervals (LCI, UCI) do not overlap, as e.g., with PIP2, PIP3, and PIP4.

The wide spreading of the presented results is caused by the individual physiological variation in vascular or skin properties, which are not completely eliminated by the proposed normalization procedure. The normalization areas are placed on finger nails, which we assume not to be strongly age dependent or having strongly varying tissue properties below the nail. The difference in age and different thicknesses of subcutaneous adipose tissue¹⁸ between volunteers and patients, which causes different optical properties, are not yet taken into account.

ROC curves have been plotted and the area under the AUC was calculated in order to quantify the discriminating power. The results are presented in Fig. 6. The best discrimination has been found for the PIP joints of the middle finger with an AUC of 0.81. Generally, normalized cumulative frequencies of normalized variances show an improved discrimination power compared to raw data analysis (see Figs. 3, 4).