The Topography of Striatal Dopamine and Symptoms in Psychosis: An Integrative Positron Emission Tomography and Magnetic Resonance Imaging Study

Background Striatal dopamine dysfunction is thought to underlie symptoms in psychosis, yet it remains unclear how a single neurotransmitter could cause the diverse presentations that are observed clinically. One hypothesis is that the consequences of aberrant dopamine signaling vary depending on where within the striatum the dysfunction occurs. Positron emission tomography allows for the quantification of dopamine function across the striatum. In the current study, we used a novel method to investigate the relationship between spatial variability in dopamine synthesis capacity and psychotic symptoms. Methods We used a multimodal imaging approach combining 18F-DOPA positron emission tomography and resting-state magnetic resonance imaging in 29 patients with first-episode psychosis and 21 healthy control subjects. In each participant, resting-state functional connectivity maps were used to quantify the functional connectivity of each striatal voxel to well-established cortical networks. Network-specific striatal dopamine synthesis capacity (Kicer) was then calculated for the resulting connectivity-defined parcellations. Results The connectivity-defined parcellations generated Kicer values with equivalent reliability, and significantly greater orthogonality compared with standard anatomical parcellation methods. As a result, dopamine-symptom associations were significantly different from one another for different subdivisions, whereas no unique subdivision relationships were found when using an anatomical parcellation. In particular, dopamine function within striatal areas connected to the default mode network was strongly associated with negative symptoms (p < .001). Conclusions These findings suggest that individual differences in the topography of dopamine dysfunction within the striatum contribute to shaping psychotic symptomatology. Further validation of the novel approach in future studies is necessary.


18F-DOPA PET Data Acquisition and Analysis
Participants were not permitted to smoke or consume caffeine for four hours preceding the scan. After acquiring a CT scan for attenuation correction, PET images were acquired using a Siemens Biograph HiRez XVI PET scanner (Siemens Healthcare, Erlangen, Germany) at Imanova Centre for Imaging Sciences.
One hour prior to scanning, participants received 400mg entacapone and 150mg carbidopa, to prevent formation of radiolabelled metabolites and reduce peripheral metabolism.
Approximately 160 MBq of 18 F-DOPA was administered by bolus intravenous injection. The quantification pipeline was consistent with previous works. 3 Correction for head movement during the scan was performed by denoising the non-attenuation-corrected dynamic images using a level 2, order 64 Battle-Lemarie wavelet filter. Frames were realigned to a single reference frame, acquired 20 minutes post-injection, employing a mutual information algorithm. 4,5 The transformation parameters were then applied to the corresponding attenuated-corrected dynamic images, creating a movement-corrected dynamic image, which was used in the analysis. Realigned frames were then summated to create an individual motion-corrected reference map for the brain tissue segmentation. The cerebellum was used as a reference region (defined as per Hammers et al 6 ), and Ki cer was calculated with the Patlak-Gjedde graphical approach adapted for reference tissue input function. 7 Image processing and quantification was done using in-house code with MATLAB 2012b, and SPM8 (Wellcome Trust Centre for Neuroimaging) was used to automatically normalize a tracer-specific template. 8 In order to generate the voxelwise Ki maps we implemented a previously established method 9 in which Ki cer parametric images of the brain were constructed from motion-corrected images using a wavelet-based approach. 10

MRI Acquisition and Preprocessing
Participants also received a rfMRI scan on a 3T GE Signa MR scanner. Image pre-processing was performed using a standard pipeline implemented in the CONN toolbox (version 17.b) 11 for Statistical Parametric Mapping software (SPM 12 (6906)). A standard preprocessing pipeline was used consisting of slice timing correction, realignment, and normalisation to MNI space. Images were smoothed with a Gaussian kernel of 8mm fullwidth-half-maximum. The ART toolbox was used to account for motion and artefact detection using anatomical component based correction (aCompCor) of temporal confounds relating to head movement and physiological noise. This method models noise effects at a voxel level based on estimates derived from principal components of noise regions of interest (white matter and CSF, eroded by one voxel to minimise partial volume effects), and then removes these from the BOLD timeseries using linear regression. Six residual head motion parameters and their first order temporal derivatives were also entered as regressors into the first level model.
A confounding effect accounting for magnetisation stabilisation, and its first order derivative was entered. A scrubbing procedure was also implemented in that artifact/outlier volumes (average intensity deviated more than 5 standard deviations from the mean intensity in the session, or composite head movement exceeded 0.9 mm from the previous image) were also regressed out. Preprocessed data were temporally bandpass filtered (0.008-0.09 Hz)

Cortical Network Assignment
Time series were extracted from the 333 nodes of the Gordon atlas. 12  We then ran the Louvain community detection algorithm on the whole brain group level graph, 13 in order to assign individual nodes to networks based on the connectivity patterns present in the current dataset. Due to the non-deterministic nature of the Louvain algorithm, a previously described consensus clustering approach was employed. 14,15 Edges between nodes closer than 10mm were discarded (euclidean minimum distance between two closest points of the two nodes), and the strongest 5% of edges were retained and binarised. The gamma parameter was set to 1.4. Networks defined by this approach were then labelled according to whichever of the apriori defined networks they showed the greatest overlap with. This identified the default mode, sensorimotor, cinguloopercular, dorsal attention, auditory and visual networks (see Figure S1). The visual network was excluded in subsequent analysis given its relative lack of direct connections with the striatum. 17

Striatal Probabilistic Parcellation
For each striatal voxel the z-transformed correlation coefficient between the voxel and the 333 Gordon nodes was calculated. Then for each of the networks defined in the Network Assignment step above the mean connectivity of that network's nodes to the voxel was calculated, with negative values being set at 0. When this had been performed for each network these values were then scaled so that at each voxel the sum of the five connectivity values (one for each network) equalled 1 (see Figure S2).

Figure S3 Striatal Probabilistic Connectivity Maps
Each striatal voxel is assigned a value for each of the cortical networks, based on the mean connectivity of the voxel to each node in the network. The total connectivity score for each voxel sums to 1. In the example above the left hand voxel shows is weighted most strongly for network 1, and least for network 3.

PET-MRI Integration
At the individual participant level for each network, for each striatal voxel we multiplied the Ki cer value at that voxel by the weight assigned to that network at that voxel. We then averaged across voxels to generate for each network a network specific Ki cer .

Reliability Analysis
Test-retest resting state MRI data was obtained from the Human Connectome project. 18 We removed 40% of the volumes from these scans so that the scan length matched that of our own data. For details of the test-retest PET data please see the previously published report. 19 Striatal parcels were defined in the same manner as for the main dataset, normalisation of these MNI parcels into individual patient space was then performed as described in the original test-retest paper 19 .

Spatial Distribution Analysis
In order to determine if patients and controls differed in terms of the spatial distribution of the connectivity based subdivisions we performed an analysis as illustrated in Figure S3.

Figure S4 Spatial Distribution Analysis
Example of quantifying the extent to which a striatal map is weighted along the x-axis. The voxel intensities of the striatum are multiplied (entrywise product) by a matrix linearly increasing along the x-axis, and the result is summed. Striatum 1 is weighted towards the right, and so scores higher than striatum 2 that is weighted towards the left.