RT Journal Article
SR Electronic
T1 Robust supervised clustering analysis of 18F DPA-714 dynamic clinical PET data
JF Journal of Nuclear Medicine
JO J Nucl Med
FD Society of Nuclear Medicine
SP 289
OP 289
VO 58
IS supplement 1
A1 Love, Scott
A1 Jaouen, Vincent
A1 Annan, Mariam
A1 Cottier, Jean-Philippe
A1 Vercouillie, Johnny
A1 Guilloteau, Denis
A1 Santiago-Ribeiro, Maria-Joao
A1 Tauber, Clovis
YR 2017
UL http://jnm.snmjournals.org/content/58/supplement_1/289.abstract
AB 289Objectives: In the healthy brain there is minimal expression of the 18 kDa translocator protein (TSPO); however, it is up-regulated due to neuroinflammation caused by, for example, cerebral stroke1. Such clinical changes in the density of this TSPO can potentially be measured using the [18F]DPA-714 radioligand and dynamic positron emission tomography2,3 (PET). To avoid the invasive collection of arterial blood sampling, which is required for the gold standard analysis of dynamic PET studies4 (2T4k) the creation of alternative reference kinetics are required to enable the use of simplified quantification techniques. Recent [18F]DPA-714 studies have used an anatomically defined cerebellum as reference region for quantification5,6; however, this approach may not be robust nor invariant for some populations of subjects. For the kinetic modelling of the [11C]-(R)-PK11195 molecular marker a supervised clustering approach (SVCA) was proposed7, and modified8, as a more robust method of creating reference kinetics. Here we aimed to improve the SVCA method and to test its ability to robustly define reference regions for use in the quantification of clinical [18F]DPA-714 dynamic PET studies.Methods: Dynamic PET scans (59 min) were acquired from 2 stroke patients and 8 healthy subjects using a Philips Ingenuity TF camera. Dynamic data were binned into 31 time frames: 6x10 sec, 8x30 sec, 4x60 sec, 5x120 sec, 8x300 sec. Structural T1-weighted MRI were acquired from all participants and rigidly registered to the mean image of the dynamic PET data using Freesurfer. During definition of the SVCA predefined classes, dynamic PET data were normalized (mean activity in brain mask subtracted from the activity of each voxel and divided by the standard deviation) to scale each frame. The predefined kinetic classes were calculated as the mean, across voxels and participants, of the normalized PET data from regions defined in an iterative process as containing blood, white matter, and low specific binding grey matter using the T1 image and Freesurfer. The specific binding grey matter class was defined by expert manual segmentation of hyperintense voxels in the reconstructed PET data of stroke patients. Nonnegative least-squares linear regression was used to quantify the contribution of each of the 4 classes to the kinetic of each voxel. For each subject, a reference curve was created as the weighted combination of the time activity curves (TAC) of a selection of voxels. The selection was based on percentiles of the weights of their decomposition over the 4 classes. An SRTM2 approach was used to calculate voxel-wise binding potentials (BP), both using the TAC of the cerebellum and the generated curve as references.Results: Using largely automated techniques, coherent TACs were created for all four classes: nonspecific grey matter, white matter, blood and TSPO. Importantly, stroke damaged brain regions were successfully detected using the proposed version of SVCA to define the reference region required for the SRTM2 compartment model. From area under the curve values, there was no significant difference [t(18)=.09,p=0.9] between the cerebellum and SVCA generated reference TACs. There were significant correlations between the BPs within a region of interest (e.g., thalamus) calculated either using the SVCA or cerebellum references.Conclusion: The modified SVCA method presented here can be used with [18F]DPA-714 dynamic PET data to detect brain tissue with neuroinflammation caused by stroke. In this study where the subjects were supposedly devoid of neuroinflammation in the cerebellum, the reference curves found by the proposed approach were very similar to those of the cerebellum while not being based on any anatomical prior. Future work should continue to optimize the method and to investigate the generalizability of the current finding to other clinical cases where neuroinflammation can occur significantly in the cerebellum. Research Support: NICAD ANR-13-PRTS-0021 $$graphic_727FF98E-D554-4484-B7C6-49896826E7D8$$