Philippe Gervais

Machine Learning for Neuroimaging with Scikit-Learn

Statistical machine learning methods are increasingly used for neuroimaging data analysis. Their main virtue is their ability to model high-dimensional datasets, e.g., multivariate analysis of activation images or resting-state time series. Supervised learning is typically used in decoding or encoding settings to relate brain images to behavioral or clinical observations, while unsupervised learning can uncover hidden structures in sets of images (e.g., resting state functional MRI) or find sub-populations in large cohorts. By considering different functional neuroimaging applications, we illustrate how scikit-learn, a Python machine learning library, can be used to perform some key analysis steps. Scikit-learn contains a very large set of statistical learning algorithms, both supervised and unsupervised, and its application to neuroimaging data provides a versatile tool to study the brain.

Early and protective microglial activation in Alzheimer's disease: a prospective study using 18F-DPA-714 PET imaging.

While emerging evidence suggests that neuroinflammation plays a crucial role in Alzheimers disease, the impact of the microglia response in Alzheimers disease remains a matter of debate. We aimed to study microglial activation in early Alzheimers disease and its impact on clinical progression using a second-generation 18-kDa translocator protein positron emission tomography radiotracer together with amyloid imaging using Pittsburgh compound B positron emission tomography. We enrolled 96 subjects, 64 patients with Alzheimers disease and 32 controls, from the IMABio3 study, who had both (11)C-Pittsburgh compound B and (18)F-DPA-714 positron emission tomography imaging. Patients with Alzheimers disease were classified as prodromal Alzheimers disease (n = 38) and Alzheimers disease dementia (n = 26). Translocator protein-binding was measured using a simple ratio method with cerebellar grey matter as reference tissue, taking into account regional atrophy. Images were analysed at the regional (volume of interest) and at the voxel level. Translocator protein genotyping allowed the classification of all subjects in high, mixed and low affinity binders. Thirty high+mixed affinity binders patients with Alzheimers disease were dichotomized into slow decliners (n = 10) or fast decliners (n = 20) after 2 years of follow-up. All patients with Alzheimers disease had an amyloid positive Pittsburgh compound B positron emission tomography. Among controls, eight had positive amyloid scans (n = 6 high+mixed affinity binders), defined as amyloidosis controls, and were analysed separately. By both volumes of interest and voxel-wise comparison, 18-kDa translocator protein-binding was higher in high affinity binders, mixed affinity binders and high+mixed affinity binders Alzheimers disease groups compared to controls, especially at the prodromal stage, involving the temporo-parietal cortex. Translocator protein-binding was positively correlated with Mini-Mental State Examination scores and grey matter volume, as well as with Pittsburgh compound B binding. Amyloidosis controls displayed higher translocator protein-binding than controls, especially in the frontal cortex. We found higher translocator protein-binding in slow decliners than fast decliners, with no difference in Pittsburgh compound B binding. Microglial activation appears at the prodromal and possibly at the preclinical stage of Alzheimers disease, and seems to play a protective role in the clinical progression of the disease at these early stages. The extent of microglial activation appears to differ between patients, and could explain the overlap in translocator protein binding values between patients with Alzheimers disease and amyloidosis controls.

[18F]DPA-714, [18F]PBR111 and [18F]FEDAA1106-selective radioligands for imaging TSPO 18 kDa with PET: automated radiosynthesis on a TRACERLAb FX-FN synthesizer and quality controls.

Imaging of TSPO 18 kDa with PET is more and more considered as a relevant biomarker of inflammation in numerous diseases. Development of new radiotracers for TSPO 18 kDa has seen acceleration in the last years and the challenge today is to make available large amounts of such a radiotracer in compliance with GMP standards for application in humans. We present in this technical note automated productions of [(18)F]DPA-714, [(18)F]PBR111 and [(18)F]FEDAA1106, three promising radiotracers for TSPO 18 kDa imaging, using a TRACERlab FX-FN synthesizer. This note also includes the quality control data of the validation batches for the manufacturing qualification of clinical production of [(18)F]DPA-714.

18F-FDG-PET patterns of surgical success and failure in mesial temporal lobe epilepsy

Objective: To search for [18F]-fluorodeoxyglucose (FDG)-PET patterns predictive of long-term prognosis in surgery for drug-resistant mesial temporal lobe epilepsy (MTLE) due to hippocampal sclerosis (HS). Methods: We analyzed metabolic data with [18F]-FDG-PET in 97 patients with MTLE (53 female participants; age range 15–56 years) with unilateral HS (50 left) and compared the metabolic patterns, electroclinical features, and structural atrophy on MRI in patients with the best outcome after anteromesial temporal resection (Engel class IA, completely seizure-free) to those with a non-IA outcome, including suboptimal outcome and failure. Imaging processing was performed with statistical parametric mapping (SPM5). Results: With a mean follow-up of >6 years (range 2–14 years), 85% of patients achieved a class I outcome, including 45% in class IA. Class IA outcome was associated with a focal anteromesial temporal hypometabolism, whereas non-IA outcome correlated with extratemporal metabolic changes that differed according to the lateralization: ipsilateral mesial frontal and perisylvian hypometabolism in right HS and contralateral fronto-insular hypometabolism and posterior white matter hypermetabolism in left HS. Suboptimal outcome presented a metabolic pattern similar to the best outcome but with a larger involvement of extratemporal areas, including the contralateral side in left HS. Failure was characterized by a mild temporal involvement sparing the hippocampus and relatively high extratemporal hypometabolism on both sides. These findings were concordant with electroclinical features reflecting the organization of the epileptogenic zone but were independent of the structural abnormalities detected on MRI. Conclusions: [18F]-FDG-PET patterns help refine the prognostic factors in MTLE and should be implemented in predictive models for epilepsy surgery.

TAU PET IMAGING IN PATIENTS WITH PROGRESSIVE AMNESIA NOT DUE TO AD

(Table 1) scanned with F-AV-1451 PET was evaluated. All tau PET SUVR images were qualitatively interpreted using 4 rating scales. Three rating scales visually assessed F-AV-1451 uptake in 7 cortical regions (medial, inferior and lateral temporal, medial and lateral parietal, frontal and occipital) either by presence/ absence of uptake or on a severity scale, from 0 to 2 and 0 to 3. The fourth rating scale looked at F-AV-1451 uptake in 3 regions corresponding to the Braak stages (Braak I/II, III/IVand V/VI). The results of the visual interpretation were compared to the average global F-AV-1451 SUVR as well as amyloid status and clinical diagnosis. The rating scale providing best correlations with global tau quantification was chosen for further analysis. A second reader independently interpreted the 60 scans and inter-rater agreement was assessed. Results: Final scores of all visual rating scales were significantly and strongly associated with quantification of global F-AV-1451 uptake (Table 2). The strongest correlation was found with the scale allowing differentiation between mild, moderate and intense uptake across 7 regions (correlation coefficient 1⁄4 0.92). As expected, mean visual rating scores within each diagnostic category were lower for amyloid negative subjects than for amyloid positive subjects. Moreover, we found an increasing trend directly correlated with diagnosis, with controls having the lowest, MCI an intermediate, and AD the highest values (Figure 1 and Table 3). Two-way ANOVA model confirmed that both clinical diagnosis (p<0.001) and PIB status (p<0.001) independently contributed to the visual score. Interrater agreement was excellent (linear weighted Kappa 0.833; correlation coefficient 0.946; R 0.915 – Figure 2). Conclusions: Our data show that visual rating scales strongly correlate with tau SUVR quantification, particularly when allowing differentiation between mild, moderate and intense uptake. This is a reliable, easily reproducible and promising alternative approach to tau measurement in clinical settings.

PROGRESSION OF CORTICAL MICROGLIAL ACTIVATION IN ALZHEIMER’S DISEASE: A TWO-YEAR LONGITUDINAL PET STUDY USING [ 18 F]DPA-714

Lorraine Hamelin, Julien Lagarde, Guillaume Doroth ee, MarieClaude Potier, H el ene Corne, Ludovic Fillon, Fabien Caill e, Philippe Gervais, Michel Bottlaender, Marie Sarazin, Neurology of Memory and Language, Universit e Paris Descartes, Sorbonne Paris Cit e, INSERM UMR S894, Centre Hospitalier Sainte Anne, Paris, France; INSERM-UPMC UMRS 938, Centre de Recherche Saint-Antoine, Paris, France; Brain and Spine Institute (ICM), UPMC/Inserm UMR-S 975, CNRS UMR 7225, Paris, France; Sorbonne Universit es, Universit e Pierre et Marie Curie (UPMC), Paris, France; CATI Project, Paris, France; CEA, Orsay, France; NeuroSpin, Institut d’Imagerie Biom edicale, Direction des Sciences du Vivant, Commissariat a l’Energie Atomique, Gif sur Yvette, France; Neurologie de la M emoire et du Langage, Universit e Paris Descartes, Sorbonne Paris Cit e, INSERM UMR S894, Centre Hospitalier Sainte Anne, Paris, France. Contact e-mail: J.LAGARDE@ ch-sainte-anne.fr