Parkinson's disease (PD) has a lifetime risk of about 2%, making it the second most common neurodegenerative disease after Alzheimer's disease. A main risk factor for PD is aging and current demographic trends predict a doubling in the number of cases by 2,050 (Bach et al., 2011). Clinically, the syndrome presents with the association of motor slowness (hypokinesia), with muscle rigidity and/or tremor and/or a postural instability (Greenberg et al., 2012). One of the neuropathological hallmarks of PD is a death of dopaminergic neurons in the substantia nigra (SN) pars compacta. Neuronal loss begins in the lateral ventral tier of the SN and throughout the illness this remains the most severely affected region. Post-mortem studies have shown that loss of cells from the SN in PD results mainly in profound dopamine (DA) depletion in the motor region of the striatum, nigral projections to the dorsal, and caudal putamen being most affected.

Because a biochemical hallmark of PD is a deficiency of striatal DA, most imaging studies have focused on studying the problem directly, using a variety of methods (Antonini et al., 1997; Constantinescu et al., 2011; Garraux et al., 2013). Measurement of striatal ¹⁸F-DOPA uptake with PET is still regarded by many as the “gold standard” for the diagnosis of PD. Several studies have reported that patients with sporadic PD have lost up to 50% of normal ¹⁸F-DOPA uptake from the caudal putamen contralateral to the side with the most severe symptoms compared with 20–30% on the ipsilateral side. The pre-synaptic striatal dopamine deficiency state can be assessed in vivo using other nuclear medicine techniques targeted to dopamine transporters (DaT) such as ¹²³I-ioflupane (also known by its tradename DaTSCAN) SPECT. These techniques are able to image a pre-synaptic striatal dopamine deficiency state shared by PD and atypical parkinsonian syndromes (APS) such as multiple system atrophy (MSA) (Wenning et al., 2013) and progressive supranuclear palsy (PSP) (Williams and Lees, 2009). They can be used to assist the clinicians to distinguish between these conditions and normal controls, essential and dystonic tremors, drug-induced, and psychogenic parkinsonism. However, it is generally considered that imaging pre-synaptic striatal dopaminergic deficit does not reliably discriminate between PD and APS.

In comparison with PD, the latter are also characterized by a post-synaptic striatal impairment. This can be demonstrated using two main approaches. The first involves ¹⁸F-FDG PET, which allows assessing the pattern of resting glucose metabolism throughout the brain. Several studies from independent groups consistently showed that FDG uptake pattern in the striatum and other regions can accurately discriminate APS from PD (Ghaemi et al., 2002; Garraux et al., 2013). These studies also showed that ¹⁸F-FDG PET may also provide some clinically-relevant information for the distinction between APS subgroups. A post-synaptic striatal deficiency state can be specifically demonstrated using radioligands of D_2/3 striatal dopamine receptors. In PD, ¹¹C-Raclopride PET studies showed preserved (or even mildly increased in de novo cases) striatal binding. In contrast, striatal ¹¹C-Raclopride D₂ binding is reduced in APS although significant decreases might only present in around 50% of individuals. Aside from ¹¹C-Raclopride, ¹²³I-Iodobenzamide (¹²³I-IBZM) SPECT has also been used to image post-synaptic striatal dopamine D_2/3 receptors. However, a meta-analysis of IBZM-SPECT studies showed that its negative predictive value was low (Vlaar et al., 2007), while another study suggest that its diagnostic accuracy was considerably lower than that of ¹⁸F-FDG PET (Hellwig et al., 2012). Finally, studies showed that ¹¹C-Raclopride PET and ¹⁸F-FDG-PET have comparable discrimination accuracy performance for the distinction between PD and MSA (Antonini et al., 1997). One of the drawbacks of ¹¹C-Raclopride is its limited clinical availability due to the short physical half-life of ¹¹C (20 min), what requires to have a local cyclotron or radiochemistry unit in the neuroimaging center. This impediment to the wider use of PET for clinical dopamine receptor studies could be overcome through the use of suitable ¹⁸F-labeled radioligands (la Fougère et al., 2010). Thus, ¹⁸F-desmethoxyfallypride (¹⁸F-DMFP) has recently emerged as a possible alternative to ¹¹C-Raclopride. Its relatively long physical half-life (109 min) enables clinical applications for differentiating between PD and APS. A study found that differences in striatal binding in the posterior putamen between PD and APS provided the larger diagnostic performance using this technique (la Fougère et al., 2010). Furthermore, as compared with IBZM and ¹¹C-Raclopride, ¹⁸F-DMFP enables D_2/3 imaging in extrastriatal regions but this property has not been exploited for the distinction between PD and APS.

Neuroimaging data used to assist the diagnosis of parkinsonism is often analyzed by means of proprietary software to delimit regions of interest (ROIs) and quantify the radiopharmaceutical uptake (Morton et al., 2005; Tossici-Bolt et al., 2006). More sophisticated systems based on machine learning were proposed for FDG PET (Garraux et al., 2013) and DaTSCAN (Prashanth et al., 2014). For the latter modality, several machine learning approaches were presented. For example, in Illán et al. (2012) the voxel intensities of DaTSCAN neuroimages were used as features along with several classifiers to separate PD patients and controls subjects. In Towey et al. (2011) and Segovia et al. (2012) two methods based, respectively, on Principal Component Analysis and Partial Least Squares were proposed to extract relevant features from DaTSCAN data. Structural data were also used to assist the diagnosis of PD (Rana et al., 2015), including the separation of PD and APS (Salvatore et al., 2014). However, the validity of DMFP data to feed statistical classification procedures is still poorly covered. The intensity profile of this imaging modality is meaningfully different of those for DaTSCAN or FDG. In the latter case, the intensity is distributed over a large number of regions, conversely for DaTSCAN most of the signal is gathered in the striatum. DMFP is half-way between DaTSCAN and FDG in terms of its intensity profile: a large proportion of the activity is located in the striatum but other regions also have intensity values large enough to contain patterns allowing the inter-subject discrimination.

In this work, we present a model to automatically distinguish between idiopathic parkinsonism and the non-idiopathic variants using DMFP data. It is based on the application of a machine learning algorithm that consider several regions of interest by means of a multiple kernel approach. We evaluated the proposed model for both, the binary separation of idiopathic and non-idiopathic parkinsonsim and the multigroup separation of PD, MSA, and PSP patients. In addition, we studied the usefulness of combining DMFP and DaTSCAN data in a single computer system. This work is intended to be the baseline of multivariate analyses of DMFP data.

Several experiments were conducted in order to evaluate the application of the multivariate approach described above to our DMFP dataset. Initially, we address the binary diagnosis problem, i.e., the separation of two groups: idiopathic and non-idiopathic (includes MSA and PSP disorders) parkinsonism. Table 2 shows the classification measures obtained in this case and compares the results obtained by the proposed MKL approach with the ones achieved by classical approaches such as using only the voxels at the striatum or using all the voxels inside the brain. The accuracy, sensitivity and specificity of the systems were estimated using a k-fold (k = 10) cross-validation scheme. All the systems were based on SVM classification and the parameters (trade-off parameter C and kernel weights q_m for MKL approaches) were estimated using a grid search (Varma and Simon, 2006). Values C = 2^e, e = {−3, −2, …, 5} and q_m = {0.1, 0.2, …0.9} satisfying ∑m=1Nkqm=1 were assessed.

In order to avoid biased results, the selection of the regions used in the MKL approach was performed using only the training dataset, i.e., in each iteration of the cross-validation loop, we selected the five regions with the most weight (note that the striatum is divided in two regions in the AAL atlas: putamen and caudate nucleus). The putamen, caudate nucleus, olfactory bulb, and thalamus were among the five selected regions in all the iterations. The supplementary motor area was the fifth region in 92% of all cases. The pseudo-code corresponding to this experiment is shown in Algorithm 1.

As shown in Table 2, the proposed method based on multiple kernel learning achieved an accuracy rate of 73.56%. The statistical significance of this measure was assessed by means of a permutation test. In this procedure, the classification algorithm was run 1,000 times using different label sets (generated as random permutations of the original label set). A p-value was then calculated as the fraction of the executions in which the accuracy was greater than or equal to the accuracy observed when using the correct labels (Pereira et al., 2009). As a result, a p-value of 0.001 was obtained. Figure 4 shows the histogram corresponding to the accuracy rates obtained in the procedure.

Subsequently, the tree-groups separation problem was addressed. In this case, the system was trained to differentiate between the disorders considered in our study: PD, MSA, and PSP. To this end, an ensemble of SVM classifiers and the one-against-one strategy was used. The results are gathered in Table 3.

In both classification approaches, binary and multiclass, the results obtained by using DMFP data were compared with the ones obtained by DaTSCAN neuroimages. For the latter neuroimage modality only the voxels at the striatum were considered as it is common in the field. Additionally, an experiment that combined both neuroimages modalities (DaTSCAN and DMFP) was performed. It allows us to evaluate if the information contained in DaTSCAN neuroimages can complement that contained in DMFP data. To this end, we extend the MKL approach to include six regions: the five regions selected for DMFP data and striatum from the DaTSCAN neuroimages. The results are included in Tables 2, 3.

Separating idiopathic and non-idiopathic parkinsonian patients in an early stage is a challenge because both groups can show similar clinical signs and symptoms (Litvan, 1999). In addition, neuroimaging techniques such as DaTSCAN, widely used to assist the diagnosis of PD, may have difficulty to differentiate between PD and MSA or PSP, as shown in our experiments. In this work, we evaluated the use of ¹⁸F-DMFP PET in computer systems to distinguish between three parkinsonian syndromes (PD, MSA, and PSP), yielding accuracy rates about 70%. Using a multivariate analysis, we found that the striatum, olfactory bulb, thalamus, and supplementary motor area are the most important regions to separate the groups. These findings were partly corroborated through univariate analysis and are consistent with previous results reported in recent studies. For example, in Chen et al. (2014) the authors found that the volumes of the olfactory bulb and tract were significantly reduced in idiopathic PD patients compared to normal subjects and MSA patients. This corroborate the idea that the mechanisms that produce idiopathic PD also affect this region.

This result obtained by hypothesis free analysis of DMFP data is in line with clinical and histopathological observations. Clinically, olfactory testing can be used to identify patients in the premotor phase of Parkinson's disease (Berardelli et al., 2013). The olfactory deficits is caused by PD pathology in the olfactory bulb and can already be observed stage I of the Braak's staging of PD pathology (Braak et al., 2002). The key role of striatal pathology in MSA and PSP is widely established and also reflected in our hypothesis free approach (Litvan et al., 1997). On the other hand, evidences of the involvement of the motor cortex in PD were reported in Vacherot et al. (2010); Lindenbach and Bishop (2013), although results have been controversial (Viaro et al., 2011). Finally, the relationship between PD and thalamus is not yet proven on pathological bases and therefore might reflect functional disturbances on basal ganglia circuits that involve the thalamus. In short, our experiments corroborate that there exist differences in those regions (olfactory bulb, supplementary motor area, and thalamus) between PD, MSA, and PSP patients.

In this manuscript, we propose a methodology to analyze DMFP data for diagnostic purposes. The primary idea is to use not only the striatum region but additionally to take into account the other regions of the brain. In order to address the small sample size problem and overcome the differences in terms of size between different regions, we propose to use a MKL approach, which is considered to be a good solution to combine heterogeneous data sources into one classification procedure (Gönen and Alpaydın, 2011; Segovia et al., 2014). In some sense, a MKL classification procedure as the one used in this work can be seen as a kind of regularization process that assigns the same weigh to all the voxels in a specific region instead of a different weigh per each voxel as a standard SVM classifier does.

Two classification approaches were addressed. The first one consisted on separating idiopathic and non-idiopathic parkinsonism (binary classification) whereas the second one was able to differentiate between PD, MSA, and PSP data (multiclass classification). The proposed methodology based on MKL classification yields higher accuracy rates than classical approaches in both cases, however these results are limited by the relative small number of patients included in this study. Unfortunately the availability of these kind of data, especially the ones acquired using new or uncommon radiopharmaceuticals is limited. The significance of the accuracy rate achieved by the proposed method was estimated by means of permutation tests obtaining a p = 0.001. Recently, Noirhome et al. demonstrated that permutation tests are more adequate than classical approaches as binomial assessment to calculate the significance of classification measures estimated using cross-validation schemes (Noirhomme et al., 2014). It is worth noting the relatively low accuracy rates obtained for MSA and PSP classes in the multiclass approach. These results suggest that the differences between these disorders do not appear in DMFP data. This is not unique to this data modality. Previous studies based on FDG data (Garraux et al., 2013) also faced the same issue. Finally, we included an experiment combining DMFP and DaTSCAN data in the same system. The results showed that including DaTSCAN data helps in the separation problem when only the striatum voxels were used however, when the additional DMFP regions proposed in this work (the putamen, caudate nucleus, olfactory bulb, thalamus, and supplementary motor area) were included (in the MKL-based approach) DaTSCAN data did not increase the separation ability of the classifier. All in all, DaTSCAN data is probably not necessary to differentiate between PD, MSA, and PSP patients if we use new systems based on DMFP data that take advantage of all the information contained in these neuroimages (not only the striatum).

Conception or design of the work: FS, JG, JR, FM, JL, GG, CP. Data collection: JL, MS, MB, AR, KB. Data analysis and interpretation: FS, JG, JR, FM, CP. Drafting the article: FS, JG, JR, FM, GG, CP. Critical revision of the article: FS, JG, JR, JL, CP.