Spinal cord hypermetabolism extends to skeletal muscle in amyotrophic lateral sclerosis: a computational approach to [18F]-fluorodeoxyglucose PET/CT images

The study included 62 patients retrospectively recruited from our published database with definite, probable, or probable laboratory-supported diagnosis of spinal-onset ALS according to the revised El-Escorial criteria [3]. None of the enrolled patients had any history of other neurological disorders, cerebrovascular disease, diabetes mellitus, or systemic inflammatory disease. All subjects provided signed informed consent to enter the study that was approved by the Ethics Committees of IRCCS Ospedale Policlinico San Martino in Genoa and of AUO Città della Salute e della Scienza in Turin, Italy.

As part of our clinical procedure, patients were submitted to our follow-up program by medical examination or phone interview. Over the follow-up period of 60 months data were available for 56/62 patients while no information could be obtained for the remaining six.

Findings obtained in ALS patients were compared with the corresponding data in control subjects selected from two different databases. Metabolic activity and structure of spinal cord as well as of psoas muscle were compared with the corresponding findings in 62 subjects selected submitted to FDG PET/CT scan > 1 year after complete removal of histologically diagnosed melanoma, subsequent histologically negative sentinel lymph node, and no evidence of relapse at least 2 years after surgery [4]. These records were extracted from the databases of the two centers and selection of each subject was performed to optimize the case-control criterion according to the used scanner, sex, and age.

For brain analysis, FDG uptake of ALS patients was compared with the published corresponding data in 44 normal volunteers with normal findings at neuropsychological evaluation and brain MRI as previously defined [5].

All PET/CT scans were acquired according to current guidelines [6, 7]. All subjects were studied in the early morning after fasting for 12 h. Serum glucose was assessed, and an antecubital vein was cannulated. Patients were invited to lie for 20 min in a silent and darkened room, with eyes closed and ears unplugged. FDG (4.8-5.2 MBq/kg body weight) was then injected 45–60 min before 3D scan using an integrated PET/CT scanner (Biograph 16-Hirez, Siemens or Discovery ST-E System, GE Healthcare). In all cases, the 15-min cerebral acquisition was followed by whole-body imaging in arms down position.

In both centers, PET data were reconstructed into a 128 × 128 matrix using a 3D iterative reconstruction algorithm (OSEM, three iterative steps, eight subsets). Raw images were scatter-corrected and processed using a 3D Gaussian filter, while CT was used for attenuation correction.

Image quality control documented a spatial resolution of 4.0 mm full width at half-maximum for both scanners. According to standard procedures of both labs, the two imaging systems were cross calibrated using a cylinder of 20 cm diameter and 20 cm length filled with a solution containing 100 MBq of ⁶⁸Ge. Images were reconstructed with the same algorithm used for the clinical protocol [6, 7]. Finally, the entire CT dataset was co-registered with the 3D PET images using commercially available software interfaces.

Muscular FDG uptake was analyzed in both psoas muscles. This site was selected because its contractile activity is minimized with patient resting in the supine position and thus in the interval between FDG injection and PET/CT acquisition; moreover, a large part of its volume is systematically included in a whole-body PET/CT acquisition and, finally, its size and structure have been proposed as relevant prognostic predictors in different disease states [8, 9].

Usually, psoas muscles are evaluated at CT by selecting a single muscle slice at the level of the third lumbar vertebra. To improve the accuracy of this evaluation, we developed a computational approach as to extract the entire recognizable muscle volume. The algorithm follows a slight modification of the general strategy previously validated by our lab for the assessment of bone marrow metabolic activity [4, 10]. According to this approach, the first step implies a visual inspection of CT images to define the proximal insertion of both psoas muscles at starting from the soma of D12 vertebra. To standardize volume definition, the caudal limit of the investigated volume was set at the plane crossing L5-S1 junction. Included slices were fed into an in-house developed software that utilizes histogram equalization and edge detection in order to segment the psoas volume. In the case of not-closed, not-connected edges, the software applies an α-shape algorithm [11] to identify the region corresponding to the inner muscle. After this automatic recognition, each slice was used to construct a binary mask, with the value set at 1 inside the domain representing the muscle and 0 elsewhere. The mask was adjusted in order to account for the differences between CT and PET pixel dimensions, downsampling the CT masks to the PET resolution. The post-processed masks were eventually multiplied against the PET data to extract the information on the FDG uptake in correspondence of the muscle voxels. The final product was thus two DICOM file series, reporting the CT and PET data, respectively. CT image was used to compute psoas volume and AAC (expressed in Hounsfield units). PET images were analyzed to estimate psoas FDG uptake, expressed as average standardized uptake value (SUV) and its heterogeneity expressed by the variation coefficient (VC-SUV, expressed in %), defined as the ratio between N-SUV SD and N-SUV average within the voxels of the two muscles of each patient. Spinal cord analysis was performed as previously described [1, 2]. Finally, myocardial FDG uptake was assessed as previously described [12]. Briefly, a volume of interest (average 6 ± 3 mL) was identified on the visible left ventricular (LV) myocardium on PET images while CT series was used as a reference, only in case of absent cardiac uptake. The myocardial volume of interest was set at a minimum value of at least 10 mL. The average SUV in this volume was estimated and divided for the corresponding average value in the liver to obtain myocardial N-SUV.

According to our procedure, all SUVs were divided by the corresponding average liver SUVs to account for possible differences in scanner sensitivity as to obtain the normalized SUVs (N-SUVs). In order to account for the obvious effect of body conformation and gender, psoas volume was normalized for the expected body volume calculated by the estimation of ideal body weight (IBW) according to the conventional formula of Robinson et al. [13].

Original DICOM data of brain acquisition were converted to NifTI-1 format using SPM8 DICOM Import [14]. PET images were normalized to a customized previously published template [15] and smoothed with an 8-mm full width at half maximum Gaussian Kernel. Brain Map Ginger ALE 2.3 (Eickhoff SB, Laird AR) was used to convert coordinates of significant clusters in the Montreal Neurological Institute (MNI) space into Talairach coordinates. Brodmann areas (BAs) were then identified at a range of 0 to 3 mm from the corrected Talairach coordinates of the SPM output isocenters, after importing the corrected coordinates by means of Talairach client (http://​www.​talairach.​org/​index.​html).

After this preliminary processing, the preprocessed NifTI-1 PET images were converted in whole-brain SUV parametric maps dividing the product between radiotracer concentration (kBq/ml) and body weight (in kg) by the injected FDG dose (in MBq) [16‐18]. Thereafter, WFU PickAtlas and NiftyReg were used to automatically identify volumes of interest corresponding to the motor cortex (Brodmann Area 4 (BA4)) in both hemispheres. All healthy control subjects were submitted to exclusive brain PET imaging. Accordingly, due to the absence of liver FDG uptake in this cohort, motor cortex FDG accumulation was analyzed considering the raw SUV since normalization for liver uptake was not possible.

All data are reported as means ± SD. Unpaired or paired t tests were used, as appropriate, to compare spinal cord N-SUV and motor cortex SUV, as well as all the computationally obtained FDG PET/CT variables describing psoas muscles (volume, AAC, N-SUV, VC-SUV). Linear regression analysis was performed using the least-squares method. A p value < 0.05 was considered significant. To assess the prognostic relevance of each of the following seven variables (i.e., age, ALS functional score and spinal cord N-SUV, psoas volume, N-SUV, VC-SUV and, finally, motor cortex SUV), the 56 patients were divided into two groups using the median value of that variable, thus resulting in two groups that differed in composition time by time according to the selected parameter. Survival was analyzed using the Kaplan–Meier method and compared using the log-rank test. Thereafter, a set of univariate and multivariate Cox proportional hazard models were fitted to the data. In the univariate analysis, the incidence of death was modeled as a function of each of the seven variables. Then, these same variables were tentatively included in a multivariate Cox model by means of a step-down (backward) procedure, based on the likelihood ratio test: variables with a p value > 0.1 were removed from the model. Proportionality assumptions were assessed as previously described [19].

Within ALS cohort, there were 36 males and 26 females, mean age was 62 ± 12 years, body weight was 69 ± 13 kg. ALSFRS-R score, collected for each patient at imaging date, ranged from 20 to 46/48 (average 39 ± 5). Overall, 22 and 40 subjects for either cohort were acquired with Siemens Hirez or GE Discovery, respectively. The time elapsed from ALS onset and PET/CT scanning was 18 ± 14 months (range 3–82). For the 56 available patients, follow-up lasted 26 ± 14 months after imaging (median 22 months, range 2–58 months). During this period, 21/56 (37%) patients died, in most cases because of respiratory complications.

In agreement with our previous reports, the spinal cord and motor cortex displayed an opposite metabolic response to ALS. As shown in Table 1, the average FDG uptake of the whole spinal cord was significantly higher in patients than in controls. This difference was evident in cervical segments while it was not appreciable in the dorsal trait. This response of spinal cord metabolism was independent from demographic and clinical variables as well as from time elapsed from diagnosis to imaging. By contrast, the effect of ALS on tracer retention in the brain cortex was the opposite. Indeed, SUV in the brain cortex was lower in patients than in control subjects in both BA4 and whole brain (Table 1).

An example of psoas muscle extraction from PET/CT images in reported in Fig. 1. ALS did not affect the CT indexes of psoas muscle composition. Indeed, AAC was remarkably similar in patients and controls (39.4 ± 8.4 vs 39.1 ± 11.3 HU, respectively, p = ns). By contrast, muscle mass was significantly affected by ALS, since volume normalized for IBW was significantly lower in patients compared to control subjects (3.6 ± 1.02 vs 4.2 ± 1.33 mL/kg, respectively, p < 0.01). Average FDG uptake in psoas muscles was significantly different between the two cohorts. Indeed, psoas muscle tracer retention was higher in ALS patients (average N-SUV 0.45 ± 0.19, median 0.43, IQR 0.19) compared to controls (average N-SUV 0.29 ± 0.09, median 0.28, IQR 0.12), respectively (p < 0.001). Moreover, ALS was also associated with increased heterogeneity of FDG accumulation, since the VC-SUV was higher in patients than in controls (VC-SUV 8 ± 4%, vs 5 ± 2%, respectively, p < 0.001). Of note, this difference did not involve the non-voluntary striated muscle represented by the myocardium, whose patients average N-SUV 1.52 ± 0.88, median 1.40 and IQR 2.12 were similar compared to controls (average N-SUV 1.73 ± 1.1, median 1.80, IQR 1.08), respectively, p = ns.

ALS induced a similar response in skeletal muscles and spinal cord metabolic activity. Indeed, as shown in Fig. 2, psoas N-SUV was directly and significantly correlated with spinal cord N-SUV in patients but not in control subjects. By contrast, psoas muscle N-SUV was independent of the corresponding metabolic index in the whole brain and BA4 average SUV (Additional file 1: Figure S1).

As a first step to evaluate their prognostic relevance, we first verified the presence of significant differences in measured indexes between the 21 patients who died during the clinical follow-up compared to the 35 survivors. As shown in Fig. 3, this analysis showed that higher mortality rates were associated with a higher FDG uptake in the spinal cord and a lower FDG accumulation in BA4. Psoas muscle showed a similar metabolic index in the two groups but higher volumes and greater homogeneity of tracer retention in survivors.

This difference was largely confirmed by the Kaplan–Meier analysis. ALS mortality rate was higher in patients with BA4 SUV lower than the median value (Fig. 4a). Conversely, N-SUV greater than the median value in the spinal cord predicted poorer overall survival (Fig. 4b). Shifting to the analysis of skeletal muscle, psoas volume was devoid of any predictive power (Fig. 4c). An increase in overall tracer retention, indexed by N-SUV, showed an appreciable, though not significant, association with a higher mortality rate (Fig. 4d), that was instead predicted by a heterogeneous tracer distribution in the muscle volume as defined by VC-SUV values over the median (Fig. 4e). These evaluations were confirmed by univariate analyses. However, the multivariate Cox regression model identified the spinal cord metabolic activation as the only independent predictive biomarker (hazard ratio 5.17, 95% confidence interval 1.74–15.38, p = 0.001), as detailed in Table 2.