Introduction

123I-meta-iodobenzyl-guanidine (123I-MIBG) planar scintigraphy is used to evaluate the noradrenergic integrity of sympathetic nerve terminals in the myocardium. As confirmed by the loss of tyrosine hydroxylase immunoreactivity associated with Lewy body diseases (LBDs) [1], low accumulation of 123I-MIBG in the myocardium can serve as a biomarker of LBDs [2,3,4,5]. Specifically, semiquantitative indices such as the heart-to-mediastinum ratio (HMR) can discriminate Lewy body diseases with high sensitivity and specificity [6,7,8,9,10,11]. Unlike the HMR, the myocardial washout rate (WR) is not frequently used in research due to errors caused by low 123I-MIBG counts in the affected myocardium [6, 7, 10, 11].

The current protocol for determination of the HMR twice is suboptimal for an accurate LBD diagnostic test. The first scan ignores the first-pass extraction of 123I-MIBG, while the second does not account for the constant loss of 123I-MIBG from the myocardium. In some circumstances, the second scan might be considered redundant [12]. In addition, the lengthy pause after the first scan could increase patient discomfort. Furthermore, the current protocol does not make full use of 123I-MIBG kinetics which can, per se, trace the pathway of norepinephrine in the nerve terminals. Thus, there is ample room for improvement of this two-scan protocol, and novel modifications could address the shortcomings in assessing the kinetics of 123I-MIBG accumulation.

Several cardiac PET studies using positron-labelled catecholamine analogues provide good references for the determination of the 123I-MIBG scan duration. Tracer equilibrium occurs 30–60 min after the injection of 11C-phenylephrine [13], 11C-meta-hydroxyephedrine [14, 15], 18F-LMI1195 [16], 18F-fluoro-hydroxyphenethylguanidines [17, 18] or 18F-labelled catecholamines [19,20,21]. Since these reports were derived from studies in primate species, we infer that long dynamic scans may not be needed for 123I-MIBG.

The life cycle of norepinephrine is highly dynamic, and injected 123I-MIBG is probably transferred back and forth continuously between intra- and extracellular spaces. However, 123I-MIBG cycling cannot be measured via static scintigraphy. The time-activity curves (TACs) of 18F-labelled dopamine during uptake and loss are different in denervated terminals and impaired vesicles [22]. Thus, the limitations of static scans of 123I-MIBG with regard to assessing nerve viability in patients with advanced LBDs may be overcome by the use of 123I-MIBG TACs.

To meet a clinical need for shorter scan protocols, we developed and optimized a novel method of dynamic planar scintigraphy (DPS) for 123I-MIBG. To address the shortcoming of the existing 123I-MIBG indices, we quantified the rates of uptake and loss of 123I-MIBG in the myocardium using DPS for a cohort of consecutively enrolled patients. Since REM sleep behaviour disorder is a strong predictor of cognitive decline and development of dementia in Parkinson’s disease [23, 24], subgroup analysis was also performed. We found that the kinetics of 123I-MIBG accumulation could be extracted from a 30-min TAC and that the new kinetic indices had comparable or better discriminatory performance for LBD patients than the existing indices, particularly when used with a machine learning classifier. The resulting improvement in diagnostic performance enhances the clinical value of 123I-MIBG scintigraphy as a biomarker of LBDs.

Materials and methods

Patient recruitment

This study was approved by the Ethics Committee of the Saitama Medical Center, Saitama Medical University. Eligible patients were evaluated for suspicion of Lewy body disease by neurology and psychiatry specialists. Informed consent was obtained from all participants. Scanning of 250 consecutive participants (mean age: 70.7, 128 men and 122 women) was performed using 123I-meta-iodobenzyl guanidine scintigraphy (MyoMIBG, FUJIFILM Toyama Chemical) from October 2017 through April 2019. Additionally, diagnoses of probable rapid-eye movement (REM) sleep behaviour disorder (pRBD) were made based on the responses to the REM Sleep Behaviour Disorder Screening Questionnaire (RBDSQ) [25] and the REM Sleep Behaviour Disorder Single-question Screen Questionnaire (RBD1Q) [26]. After reviewing the patient medical records, those with ischaemic heart disease, congestive heart failure, diabetes, or medications that could affect 123I-MIBG imaging were excluded.

Scanning protocol

123I-MIBG scans were performed with a two-detector single-photon emission computerized tomography (SPECT) camera (Discovery NM 630, GE Healthcare) equipped with extended low-energy general-purpose (ELEGP) collimators. ELEGP is optimal for 123I imaging due to its high sensitivity and low septal penetration. A 10% energy window was used on the 159-keV photopeak. Planar images were obtained using a 256 × 256 matrix. Scan 1 (30 × 2 s + 40 × 6 s + 75 × 20 s, total 30 min) was started immediately after a bolus injection of 111 MBq of 123I-MIBG. Scan 2 (3 × 300 s) and scan 3 (3 × 300 s) were started 90 and 180 min, respectively, after the injection.

Preparation for plasma input and tissue output functions

To perform kinetic analysis for DPS, we obtained decay-corrected TACs of mediastinal ROIs (mROIs) and heart ROIs (hROIs) using Smart MIBG software [27]. We fitted the mediastinal TACs (mTACs) to a three-phase exponential function with a time offset and a constant term from each peak time through 30, 105, or 195 min after the injection using MATLAB R2018b (MathWorks, Natick, MA, USA). The fitted mTACs were corrected both for 123I-MIBG binding to platelets and for metabolites in the plasma. Instead of analysing blood samples with high-performance liquid chromatography (HPLC), we used a population-based blood-to-plasma ratio (BPR) curve and a population-based metabolite correction (PBMC) curve using previously published methods [28]. We set the minimum BPR to 0.6 (haematocrit: 40%) for the first 45 s. Finally, the plasma input functions (PIFs: cps/pixel) of 123I-MIBG were obtained by multiplying the fitted mTACs by PBMC/BPR. The tissue TACs (tTACs) of 123I-MIBG were obtained by subtracting the fitted mTACs from the heart TACs (hTACs). These PIFs and tTACs were then used for the subsequent analyses.

Kinetic analysis

In order to describe the kinetics of 123I-MIBG, and alongside the conventional ratio indices, we defined three new indices for 123I-MIBG DPS, as shown in Table 1: iUp, uptake rate; iLoss, loss rate, and iNs, non-specific distribution. We used a one-tissue three-parameter model (1T3P) defined by the following equation to determine these indices:

$${\text{tTAC}}\left( t \right) = {\text{PIF}}\left( t \right) \otimes {\text{iUp}} \cdot \exp \left( { - {\text{iLoss}} \cdot t} \right) + {\text{iNs}} \cdot {\text{PIF}}\left( t \right)$$
(1)

where \(\otimes\) denotes the convolution operation. A one-tissue two-parameter model (1T2P) was defined by omitting the iNs term from Eq. 1. tTACs of different frame durations (from 1 min to 5, 10, 15, 20, 25, 30, 105, and 195 min) were fitted to both model equations. Weighted nonlinear least-squares optimization was performed with MATLAB functions with a simple weighting of each frame duration. The Akaike information criterion (AIC) [29] and the Schwarz information criterion (SIC) [30] were calculated as follows to compare the model fits:

$${\text{AIC}} = N \cdot \ln \left( {{\text{WSSR}}} \right) + 2 \cdot p$$
(2A)
$${\text{SIC}} = N \cdot \ln \left( {{\text{WSSR}}} \right) + p \cdot \ln \left( N \right)$$
(2B)

where N is the number of fitted frames, p is the number of parameters, and WSSR is the weighted sum of squared residuals. Then, we used linear and nonlinear regressions to predict HMRs of the early and delayed phases (15 and 195 min, respectively) and WR from the values of iUp and iLoss of the cohort.

Table 1 Nomenclature of the kinetic indices for myocardial 123I-MIBG scintigraphy
Full size table

Comparison of diagnostic performance

We sought to compare the classification performance of existing indices between LBD and non-LBD patients with that of our new indices, iUp/iLoss (specific distribution) and iLoss. Patients were considered unclassifiable and thus excluded from this analysis if they had an inconclusive diagnosis, or concurrent LBD and non-LBD. To quantify the diagnostic performance, we used the values of the area under the ROC curve (AUC). To test for significant differences in AUCs between the indices, we used a bootstrap test of the pROC package for R. Then, to apply a machine learning (ML) classifier that takes multiple indices, we used two support vector machines (SVMs) of the scikit-learn package for classification between LBD and non-LBD patients; one SVM employed a linear kernel, and the other, a radial basis function (RBF) kernel [31] in the space of iLoss and iUp/iLoss. The imbalance in the sample size between LBD and non-LBD patients was corrected using the synthetic minority oversampling technique (SMOTE) [32]. The patient cohort was randomly split such that 70% of the patients were used to train the SVMs and the remaining 30% were used to test them. This random splitting was performed 200 times to estimate the diagnostic odds ratios (DORs) as well as AUCs associated with each of the indices.

Comparison of 123 I-MIBG turnover among LBD subgroups

We sorted the patients with Lewy body diseases into three subgroups as follows: Parkinson’s disease (PD) with probable REM sleep behaviour disorder (pRBD), PD without pRBD, and dementia with Lewy bodies (DLB). Differences in the mean values of iUp/iLoss and in the mean values of iLoss were assessed between the PD without pRBD subgroup (as the reference group) and the other two subgroups with Dunnett’s multiple comparison test. We used GraphPad Prism 8 (GraphPad Software, San Diego, CA, USA), R 4.0.2 (R Core Team), and scikit-learn 0.23.2 for Python 3.7.6 (Python Software Foundation) as needed.

Results

Patient demographics

After excluding 42 patients who met the exclusion criteria, a total of 208 patients (106 men and 102 women) were included in the kinetic analysis (105 LBD patients, 61 non-LBD patients, and 42 patients with unclassifiable parkinsonism at the final diagnosis). The demographic profiles of the 208 patients, including age, sex, and prevalence of pRBD, are shown in Tables 2 and 3. Table 2 is based on primary diagnoses prior to 123I-MIBG, while Table 3 shows final clinical diagnoses.

Table 2 Demographics of the 208 patients subjected to pharmacokinetic analysis
Full size table
Table 3 Demographics of the 166 patients subjected to receiver operating characteristic (ROC) curve analysis
Full size table

Kinetic analysis

We first generated TACs from ROIs drawn on patient images (Additional file 1: Fig. 1). After extracting the PIFs and tTACs (Additional file 1: Fig. 2), we proceeded to fit these curves to two models with and without the third index (iNs). As shown for two representative non-LBD and LBD patients in Fig. 1, the 1T3P model provided better fitting than the 1T2P model. The kinetic results are summarized in Table 4 for 1T3P and in Table 5 for 1T2P. The information criteria (AIC and SIC) were lower with 1T3P than with 1T2P. In the 1T3P series, the lowest values (AIC: − 558.5; SIC: − 550.3) were obtained at 30 min. Truncating the TACs to less than 30 min led to negative values of iUp and iLoss, while prolonging TAC acquisition to 105 or 195 min provided no additional benefit. Using the same cohort, the mean (SD) values of early HMR, delayed HMR and WR were 1.98 (0.52), 1.98 (0.77), and 0.38 (0.30), respectively. Thus, using the data from tTACs spanning 1–30 min in the 1T3P model best describes the kinetics of 123I-MIBG DPS.

Fig. 1
figure 1

The results of nonlinear least-square fitting of the two kinetic models for two representative patients. A and B are obtained from a patient without LBD, while C and D are obtained from a patient with LBD. The tTACs were fitted with three parameters (1T3P: A, C) better than with two parameters (1T2P: B, D). See Additional file 1: Figs. 1 and 2 for the corresponding ROI TACs and PIFs

Full size image
Table 4 Kinetic results of the three-parameter model (1T3P)
Full size table
Table 5 Kinetic results of the two-parameter model (1T2P)
Full size table

We next sought to construct predictors of the existing indices to ensure follow-up and continuity. Linear regression analysis revealed that iUp/iLoss was an excellent predictor of early HMR (Fig. 2A) and delayed HMR (Fig. 2B). The scatter plot between iLoss and WR indicated a good fit with an exponential monomolecular growth model (Fig. 2C). The pharmacological half-life of trapped 123I-MIBG (0.693/iLoss) was a good linear predictor of WR (Fig. 2D). The scatter plot of iLoss and iUp/iLoss for the 208 patients and the cut-off values for each parameter are shown in Fig. 3.

Fig. 2
figure 2

The results of regressions between the existing indices and new indices of 123I-MIBG myocardial scintigraphy. Linear regressions between iUp/iLoss and the early (eHMR) (A) and delayed HMR (dHMR) (B) are shown. A nonlinear regression with an exponential monomolecular growth model between iLoss and WR (C) can be converted to a linear regression between 0.693/iLoss and WR (D)

Full size image
Fig. 3
figure 3

The scatter plot of iLoss (min−1) versus iUp/iLoss obtained from 208 patients. The shaded rectangle represents a territory based on the cut-off values for iLoss and iUp/iLoss shown in Table 6. Note that high values of iLoss were counterbalanced by iUp to some extent

Full size image

Superior diagnostic performance of 123 I-MIBG DPS

To compare the diagnostic performance, 42 patients with unclassifiable parkinsonism were excluded from the ROC analysis. The demographic characteristics of the remaining 166 patients included in this analysis (105 patients with LBDs and 61 non-LBD patients) are shown in Table 2. The results of the ROC analysis are summarized in Table 6. (The corresponding curves are shown in Fig. 4.) The best diagnostic performance was obtained by iLoss, followed by iUp/iLoss. Within short scan protocols, the AUCs of these two indices were significantly higher than that of early HMR. Using iLoss and iUp/iLoss together, representative classification results obtained with single runs of the two SVMs are shown in Fig. 5. After 200 runs, the linear SVM and RBF-SVM gave mean AUCs of 0.911 and 0.916, respectively. The mean DORs were 31.6 for the two fixed cut-off values of iLoss and iUp/iLoss, 56.4 for the linear SVM, and 57.5 for the RBF-SVM, indicating that the SVMs for the two new indices may be able to better discriminate patients with LBDs from those without.

Table 6 Results of the receiver operating characteristic (ROC) curve analysis
Full size table
Fig. 4
figure 4

Receiver operating characteristic (ROC) curves of the new (A) and current (B) indices of 123I-MIBG for identifying patients with Lewy body disease (LBD). ROC analysis was performed for 166 tests (see the demographics of the 105 LBD and 61 non-LBD patients in Table 3). Areas under the curve (AUCs) are summarized in Table 6

Full size image
Fig. 5
figure 5

Representative results of machine learning classifier models for iLoss and iUp/iLoss to discriminate patients with Lewy body disease (LBD) from those with non-LBD. After correction for class imbalance using SMOTE, 210 coordinates of iUp and iUp/iLoss, which represent 105 LBD patients (squares in red) and 105 oversampled non-LBD patients (squares and triangles in blue), are shown (A). An example portion allocated for testing (63 samples, 30%) is shown with the independent cut-off values of iLoss and iUp/iLoss (B). Decision boundaries (thick lines) and predicted probabilities (thin lines) were generated from single runs of the SVM classifier models with the linear kernel (C) and the radial basis function kernel (D)

Full size image

Comparison of 123 I-MIBG turnover among LBD subgroups

Ninety-six patients had clinically established PD. Among them, 36% (35/96) were pRBD-positive, while 64% (61/96) were pRBD-negative. Nine patients were diagnosed with DLB (Table 3). The mean (SD) estimates of iUp/iLoss were 1.17 (0.47) for pRBD-positive PD patients, 2.30 (1.67) for pRBD-negative PD patients, and 1.09 (0.68) for patients with DLB. Dunnett’s multiple comparison test revealed that the mean iUp/iLoss value of the pRBD-negative PD subgroup was significantly higher than each of the other two subgroups (P < 0.001 and 0.05; Fig. 6A). The mean (SD) estimates of iLoss were 0.0647 (0.0170) for pRBD-positive PD, 0.0557 (0.0129) for pRBD-negative PD, and 0.0683 (0.0247) for DLB. Likewise, the mean iLoss value of the pRBD-negative PD subgroup was significantly lower than each of the other two (P < 0.05; Fig. 6B). Thus, the diagnostic performances of iLoss and iUp/iLoss can potentially distinguish LBD subgroups. The mean (SD) estimates of early and delayed HMRs were 1.48 (0.16) and 1.24 (0.15) for pRBD-positive PD patients, 1.80 (0.45) and 1.64 (0.60) for pRBD-negative PD patients, and 1.43 (0.20) and 1.23 (0.21) for patients with DLB. The statistical results for HMRs were similar as shown in Fig. 6C, D.

Fig. 6
figure 6

Column scatter plots of iUp/iLoss (A), iLoss (B), early HMR (C) and delayed HMR (D) comparing the three subgroups of patients with LBDs. Asterisks denote significant differences (*: P < 0.05, **: P < 0.01, ***: P < 0.001, ****: P < 0.0001) detected by Dunnett’s multiple comparison test versus the baseline group (PD without pRBD). Horizontal bars in grey denote the mean values of the three subgroups. Note that some overestimated eHMRs (underestimated severity) in patients with RBD-negative early PD might exaggerate statistical differences

Full size image

Discussion

In this study, we sought to optimize the diagnostic procedure for detection of LBDs. 123I-MIBG TACs obtained from dynamic imaging for 30 min yielded iLoss and iUp/iLoss values that could be used to distinguish LBD and non-LBD patients with an accuracy equal to or greater than the current standard indices. Thus, iLoss could serve as a biomarker for neurodegeneration in the sympathetic nerve terminals of patients with LBDs.

We used population-based correction functions to obtain plasma input functions of 123I-MIBG. Since the plasma 123I-MIBG fraction gradually declines to 30% as previously reported [28], simple rescaling of mediastinum TACs does not suffice. The metabolite correction curves cannot be ignored when estimating the kinetic indices, as with the case of 18F-FDOPA kinetic modelling [33, 34]. The normal range of iUp/iLoss would be wider than those of HMRs, due to the presence or absence of plasma radioactivity correction for metabolites that do not penetrate the terminals.

In the 1T3P model, we set the interstitial distribution (iNs) of 123I-MIBG as a fraction of the PIF. We assumed that equilibration with 123I-MIBG in the plasma occurred within 1 min post-injection, in line with the reported equilibration time of the myocardial intensity of Gd-DTPA, an MRI extracellular contrast agent [35]. The 1T3P model stems from an 18F-FDOPA model for brain PET that quantifies the turnover according to the same principle [33]. The loss rate of 123I-MIBG in planar imaging appears comparable to that of 4D imaging; indeed, the range of iLoss was in good agreement with the normal loss rate (< 0.035) of 18F-fluorodopamine [36]. We confirmed that a 30-min scan was sufficiently long compared with the mean of 0.693/iLoss (15.2 min). Presumably, steady-state 123I-MIBG trapping was achieved in 30 min by the combination of the bolus-like delivery to the myocardium, the slow plasma radioactivity excretion (Additional file 1: Fig. 1) and the stable 123I-MIBG fraction in plasma [28]. These characteristics are common among other radiolabelled catecholamine analogues [13,14,15,16,17,18,19,20,21].

iUp/iLoss was an excellent predictor of early and delayed HMRs (Fig. 2A, B). However, iUp/iLoss is not simply a presentation of the current standard indices on a different scale; rather, it is an independent indicator of the true trapping of 123I-MIBG [34]. Indeed, the y-intercept (1.18) seen in Fig. 2A was approximately the mean iNs (0.21) plus the blood pool factor. WR could be predicted both from iLoss (Fig. 2C) and from the half-life of 123I-MIBG (0.693/iLoss) (Fig. 2D), but with limited precision. These predictions for WR are based on crude assumptions, such as ignoring 123I-MIBG metabolites (see Appendix). Thus, iLoss stands as a unique index for 123I-MIBG.

Our findings revealed that the iLoss and iUp/iLoss values derived from patients with LBDs were spread over a wide range (Figs. 3, 5), replicating a scatter plot of 18F-fluorodopamine PET [36]. In essence, deviation from the normal range could be attributed to an increase in iLoss and/or decrease in iUp. Notably, the high loss of 123I-MIBG was compensated in part by viable 123I-MIBG uptake. Thus, the apparent viability even with extremely low levels of 123I-MIBG could represent “sick but not dead” nerve terminals [37]. Unlike iUp, the fractional loss rate is (by definition) independent of 123I-MIBG trapping and does not change with the density of intact terminals. Thus, the observed increased loss rate is consistent with the level of hazard to viable terminals.

Alpha-synuclein oligomers are considered crucial in the pathogenesis of LBDs [38], as they impair the homeostasis of synaptic vesicles and membranes [39,40,41]. The leakage of catecholamines from vesicles causes the accumulation of excess toxic aldehydes in the cytosol. Ultimately, this toxicity causes the aggregation of alpha-synuclein protein. Indeed, the aldehyde metabolite from norepinephrine exacerbates nerve degeneration [42]. Although we were unable to calculate the level of 123I-MIBG in the cytosol, the loss rate of 123I-MIBG could be used as a surrogate. We presume that the excess extravesicular aldehydes are key to establishing 123I-MIBG as a biomarker of LBDs. In particular, the high loss rate of 123I-MIBG might be a symptom of the ‘vicious cycle’ that underlies the catecholaldehyde hypothesis [37, 43]. Thus, individual variability in the progression rate of LBDs might be reflected by iLoss of 123I-MIBG DPS. Unlike a high iLoss value, a low iUp/iLoss value cannot be used to discriminate between a loss of terminals due to neuron death versus the reduced function of viable nerve terminals due to alpha-synuclein oligomerization.

ROC analysis (Table 6) confirmed the good diagnostic accuracy of the early and delayed HMRs in discriminating LBDs described in previous studies [6,7,8,9,10,11]. The AUCs of both iLoss and iUp/iLoss were significantly higher than that of early HMR (P < 0.05). The highest AUC (0.903) was obtained by the value of iLoss obtained at 30 min. Furthermore, the performance of the new indices obtained by 30-min DPS was on par with that of the current indices obtained at three hours (Table 6, Fig. 4). The additional use of machine learning (ML) is attractive in that oblique or curved cut-off lines can be fixed in the coordinate space. Indeed, our results (Fig. 5) suggest that a ML classifier for multiple indices could outperform a single cut-off point for iLoss.

123I-MIBG turnover was higher in the pRBD-positive PD subgroup and the DLB subgroup than in the baseline PD subgroup as indicated by iUp/iLoss and iLoss (Fig. 6). Based on the proportionality between iUp/iLoss and HMRs, these results agreed well with earlier studies using HMRs and similar populations [3, 44,45,46,47]. Patients with pRBD-positive PD and DLB tended to have higher iLoss values, which merits further analysis. Patients with DLB have a poorer survival rate than those with PD [48], and their survival rate is influenced by frequent orthostatic hypotension [49]. However, RBD predicts motor progression in patients with PD [50]. Thus, further research is required to explore the prognostic value of 123I-MIBG iLoss with regard to distinguishing these LBD subgroups.

In addition to the cyclical nature of some diagnoses, this study has a limitation in that some cases were diagnosed with PD based on the currently used 123I-MIBG scintigraphy method. Thus, the ROC curves might be biased in favour of existing indices. We recommend that the new DPS indices be compared with delayed HMRs in future replication studies. Moreover, we had no choice but to omit inconclusive patients due to the nature of our study. Fewer difficult/borderline cases are likely to be included in the diagnostic performance analysis. Thus, the performance of all indices might be systematically overestimated. Although not tested here, we suggest that this method could be applied to diagnosis of congestive heart failure, catecholamine-induced cardiomyopathy (e.g. pheochromocytoma), and presumably treatment-related complications of anti-tumour agents.

Conclusion

We show that dynamic planar scintigraphy followed by kinetic modelling is an optimal method for using 123I-MIBG as a biomarker of LBDs, as it allows a short scan duration of 30 min and quantification of 123I-MIBG turnover in the sympathetic nerve terminals. The DPS method reduces the waiting times for patients and their family members before and after the scan while maximizing diagnostic performance. High 123I-MIBG turnover might be present in the nerve terminals of patients affected by LBDs that are still viable. If so, the high loss rate of 123I-MIBG might be a biomarker of the neurotoxicity caused by LBDs. Further studies are needed to confirm this hypothesis.