Post-Doctoral Research Visit F/M "Campagne Post-doctorant" - NEURACLE : Neuroimaging Oracle for Parkinson’s Disease modelling using Representation Learning

The project is marked by a collaborative synergy among various specialized disciplines, including Aramis (Inria team-project), the Center for NeuroImaging Research (CENIR), as well as the “Movement, Investigations and Therapeutics » (Mov'it) and the “Molecular Pathophysiology of Parkinson's Disease” teams at ICM.

This integration brings together clinical expertise, which offers a profound understanding of PD, and advanced imaging techniques like segmentation, super-resolution, and tractography.

The scientific project involves specialized artificial intelligence researchers who have extensive experience in sophisticated AI and algorithms. This expertise is crucial for representation learning for MRI brain analyses that unravel the intricate digital patient’s profiles. Efficient database management plays a key role in organizing extensive data sets effectively. The project also focuses on adapting algorithms for diverse brain MRI data. Strategic interfaces between these disciplines create a dynamic and cooperative environment, leveraging the unique strengths of each area. This concerted approach aims to reveal the complex aspects of PD progression, ensuring a comprehensive and insightful exploration of the disease.

The recruited person will be working in strong collaboration with Romain Valabregue (Mov'it) and Stephane Lehericy (Mov'it and CENIR). Romain is in charge with the management and the pre-processing of the ICEBERG dataset and Stéphane is a world-reputable expert in Parkinson’s Disease, director of CENIR. On the methodological side, the candidate will work with Aramis team, concerning the deep learning an image analysis approaches, as well as explainable AI and Bayesian paradigms.

Objective of the proposed research

This study is related to a new and extremely ambitious project to build a Neuroimaging Oracle for Parkinson’s Disease modelling using Representation Learning. This generic work, involving a unique dataset (ICEBERG) – to be extended to UK Biobank and PPI - as state-of-the art unsupervised deep learning algorithms is destinated to build an extremely rich latent-space, nourishing numerous novel discoveries based on MRI biomarkers. This project will be focusing first on Parkinson’s Disease but will be the spring for a whole new generation of studies related to neuro-degenerative diseases

Summary of the research project

Parkinson's disease (PD) is a heterogeneous disease with several phenotypic presentations variably combining motor and non-motor characteristics. Better understanding this clinical heterogeneity is an important clinical question. Patient subtyping can be defined based on clinical presentation. In this case, non-motor features emerge as key classifiers including cognitive impairment, rapid eye movement sleep behaviour disorder (RBD) and dysautonomia to identify mild, intermediate, and severe forms of the disease. This approach, however, lacks cerebral correlates. Subtyping can also be done based on the mechanism of lesion spread. Based on histological and imaging data, two potential pathways of pathological progression have been suggested: a brain-first (top-down) type, where the pathology initially arises in the cortex, possibly through the olfactory system, with secondary spread to the brainstem, and a body-first (bottom-up) type, where the pathology originates in the enteric nervous system or peripheral nervous system and then spreads to the brain. These routes are thought to model the propagation of intraneuronal α-synuclein aggregates. These two types have been reported in patients without and with RBD, respectively. The latter type of progression has also been suggested to present isolated RBD as a prodromal phenotype. However, these two models do not fully account for the heterogeneity of the disease. For example, i) RBD can occur before (in the case of iRBD) or after the onset of PD, ii) cognitive dysfunction occurs more or less early in iRBD patients, leading to dementia with Lewy bodies in the first case, iii) cognitive and olfactory disorders are frequent and early symptoms in PD patients with RBD, although the bottom-up model is the one that best explains the spread of lesions in these patients. Our overall goal is to study the heterogeneity of PD phenotypic presentation and model associated spatiotemporal brain changes based on patient-specific clinical phenotypes, genetic profiles, and MRI biomarkers using pioneering unsupervised deep learning models. We hypothesize greater individual variability than rigid top-down or bottom-up trajectories in PD and iRBD. We will develop an innovative representation learning model (vision transformers) to assess brain pathological differences by analysing the algebraic properties of MRI voxels on large longitudinal cohorts with multimodal imaging resources using data from PD (with and without RBD, olfactory disorders and/or rapid, cognitive impairment) and iRBD to study the unique pathological profile of each patient type and examine their different lesion progression trajectories, thus providing a powerful digital synthesis of patient profiles. The MRI biomarkers will include innovative super-resolved diffusion-weighted (DW)-MRI images allowing to study small brainstem and basal forebrain nuclei with much higher precision than conventional DW-MRI strategies and compare abnormalities in the bundles connecting these brain structures, creating super-resolved circuits linked to sleep, olfaction, and cognitive processes.

Background and scientific / medical question

Sporadic PD is a progressive degenerative disease of the nervous system that manifests itself clinically when the pathology has already reached an advanced stage (Braak et al., 2003). Several non-motor symptoms associated with PD often precede motor dysfunction by more than ten years (Kalia and Lang, 2015). Post-mortem diagnosis of the presymptomatic and symptomatic phases of PD is based on the identification of specific inclusion bodies, the main components of which are aggregates of a-synuclein protein (normally presynaptic), which develop as spindle or thread-like Lewy neurites in the cellular processes, and as globular Lewy bodies in the neuronal perikaryal (Braak et al., 2003). Studies suggest that misfolded a-synuclein behaves in a prion-like manner leading to cell-to-cell pathology propagation (Uchihara and Giasson, 2016). However, it remains unknown from where the initial a-synuclein aggregates originate. Braak et al. (2003) proposed an influential model for the progression of Lewy pathology in a spatiotemporal pattern with six stages: deposits initially spread from the peripheral nervous system along the vagus nerve or olfactory bulb to the medulla in a caudo-rostral pathway, causing dysautonomic and olfactory disturbances. Subsequently, they further infiltrate the brainstem, leading to sleep disturbances, and later motor symptoms when the substantia nigra is affected. Finally, the lesions progress into the limbic system, and then the neocortex, explaining the occurrence of higher cognitive dysfunctions. It has been speculated that a-synuclein inclusions initially form in nerve terminals of the enteric nervous system, and subsequently spread via autonomic connections to the dorsal motor nucleus of the vagus and intermediolateral cell columns of the sympathetic system (Braak et al., 2003; Borghammer, 2018). However, other evidence suggest that PD does not start in the enteric nervous system in all cases (Parkkinen et al., 2008). An alternative route through the olfactory system has also been hypothesised (Braak et al. 2003). The team of P. Borghammer has thus proposed a model including two subtypes: a brain-first (top-down) type, where a-synuclein pathology initially arises in the brain with secondary spreading to the peripheral nervous system; and a body-first (bottom-up) type, where the pathology originates in the enteric nervous system or peripheral nervous system and then spreads to the brain (Horsager et al. 2020). They also suggested that isolated RBD, a prodromal condition for parkinsonism, follows the body-first type, as propagating bottom-up pathology will affect the pons before reaching the substantia nigra. However, this dual model does not consider the full complexity of the pathology. For instance, RBD variably occurs before or after the onset of PD. Cognitive and olfactory dysfunctions are frequent and early symptoms in PD patients with RBD and iRBD, although the bottom-up model is the one that best explains the spread of lesions in these patients. Cholinergic, noradrenergic, and serotonergic dysfunction may also contribute to olfactory loss in PD (Doty, 2012). Other factors may thus come into play such as the genetic profile, inflammation, mitochondrial dysfunction or even the specific vulnerability of the affected nuclei (Bohnen and Postuma, 2023).

The development of other approaches capable of modelling in a more flexible manner the individual variability of subjects would allow a better understanding of the heterogeneity of the phenotypic presentation and brain lesion propagation of the disease. Such a model would rely on a robust, longitudinal stratification system to classify PD and iRBD patients into distinct subgroups based on a combination of clinical and MRI biomarkers, to identify the trajectory and severity of their pathological progression. Traditional bottom-up MRI analysis approaches (Alexander et al., 2019) consist in annotating and then detecting and segmenting relevant regions of interest (ROI), highlighted by the state of the art. The model we are planning to elaborate in this study takes advantage of the important brain MRI dataset accessible by our teams. Exploiting these images by consolidating a balanced dataset and with a minimum of biases will need to use state-of-the-art powerful unsupervised methods based on visual transformers, in the optics of a robust representation-learning. Representation learning has the advantage of a massive analysis of low-level brain MRI information (voxels) without passing through the bias-generator segmentation bottleneck. Representation learning is giving spectacular results in a wide variety of domains. No such study has yet been done in neuroimaging. Generating a semantic indexing of such a massive brain MRI dataset can generate much more than a digital twin: a digital brain-oracle, which could constitute a world-reference in the area of neurodegenerative diseases. As such, this latent space will comprise a series of rich features, enabling us to interrogate, in the latent space, this brain oracle, as a digital companion, able to provide responses to questions related (for example) to the PD propagation route, to the specificity of a given dataset, a content-based retrieval mechanism, which, enriched with relevance feedback information could lead to a new way of generating biomedical insights from this fundamental biomarker. Explainable and responsible AI will also be able to help us to support such interactive workflow, by also enabling scaling-up this model to other viewpoints in PD, as well as to the extension (in the future) to other neurodegenerative diseases.

Specific Objectives

We will model the heterogeneity and spatiotemporal brain-associated changes based on patient-specific clinical phenotypes, genetic profiles, and MRI biomarkers by using an innovative unsupervised representation learning model, never used in the neuroimaging field, on multiple large-scale longitudinal cohorts with multimodal brain MRIs. This approach allows creating a generic and refined virtual brain model of PD that will be able to capture individual patient characteristics and evolution profiles.

More specific objectives are listed below:

• to determine homogeneous groups of patients, corresponding to pathological subtypes, with severe forms, based on their clinical phenotypes and imaging profiles,
• to determine the propagation routes of pathological lesions for each group of patients with PD with and without RBD, and/or anosmia, and/or rapid cognitive decline, and iRBD.

Hypotheses

We hypothesise that:

• the use of the unsupervised representation learning model, previously never used on brain MRI, will allow us to characterize the heterogeneity of PD using clinical and MRI data,
• this model will allow us to visualize those patients with a greater individual variability than rigid descending or ascending trajectories in PD and iRBD patients.

Methods

Dataset

Available clinical data: We have available the ICEBERG (N=315) monocentric cohort from the ICM and the Parkinson’s Progression Markers Initiative (PPMI) (N=1425) cohorts and aim to access the UKBioBank (N = 153,569) cohort. These are longitudinal large-scale multi-site resources for neuroimaging including PD patients with and without RBD, olfactory disorders and/or rapid cognitive impairment. These three cohorts all include neuromelanin-sensitive MRI, T1-weighted MRI, DW-MRI and quantitative susceptibility mapping (QSM), as well as clinical and genetic data. ICEBERG includes 170 idiopathic PD patients, 58 iRBD patients and 70 healthy volunteers whereas PPMI includes 423 PD patients, 67 subjects with RBD hyposmia or dopamine deficit, 196 healthy volunteers and 294 PD patients with a genetic mutation in Parkin, PINK, LRRK2 or GBA. In the ICEBERG cohort, RBD is diagnosed with polysomnography, RBD-HK test and Epworth sleepiness scale while in the PPMI cohort, RBD is diagnosed with REM Sleep Behavior Disorder Questionnaire and Epworth sleepiness scale. Both cohorts use the UPSIT test to evaluate hyposmia/anosmia in subjects. We will harmonise as best as possible the clinical scales used in the different studies. Polygenic risk score will be calculated when available.

Imaging biomarkers will include neuromelanin-sensitive, T1-weighted, DW-MRI and QSM. Neuromelanin-sensitive MRI will be used to calculate the volume and signal changes in the SN and the LC as previously done (Gaurav et al., 2022, Garcia-Lorenzo et al., 2013). T1-weighted images will be used to extract grey matter volume and cortical thickness. QSM maps will be reconstructed as described previously (Biondetti et al. 2020, 2021). DW-MRI images will be processed using super resolution (image quality transfer) technique (coded on Matlab, images seen on MRtrix3) (Alexander et al., 2017) to increase anatomical details of brainstem and basal forebrain nuclei.

Data analysis:

Representation learning methodology:

Representation learning is favoured in machine learning for its ability to autonomously extract and learn pertinent features from large datasets, a significant advancement over traditional approaches. This methodology, as highlighted in the works of Radford et al. (2021), Caron et al. (2021), Benoit (2022), Oquab et al. (2023), and Darcet et al. (2023), eliminates the need for ROI selection, which is a standard practice in machine learning. In traditional methods, experts must identify and define the ROI for dense tasks or labels for classification to be used in the learning process, a task that can be both time-consuming and susceptible to human bias. In contrast, our strategy for advancing PD research involves several key steps in building a brain MRI foundation model, Neuracle (generic Neuroimaging Oracle), for brain research: Firstly, we train self-supervised vision-transformers (ViT), as suggested by Vaswani, A et al. (2017), Liu, Z et al. (2021), and Oquab et al. (2023), on MRI data for representation learning. This step utilizes diverse MRI datasets, including ICEBERG, PPMI, and UKBioBank, which represent various stages, severities, and symptoms of PD and iRBD, as well as healthy brains. Secondly, the model, post-training, is employed in specific PD research tasks. These tasks include lesion detection (segmentation task), progression monitoring (tracking and regression tasks), and subtype stratification (classification and clustering tasks). For each of these tasks, a specific 'head' model is applied, connected to the main representation model, which remains in a frozen state to utilise its learnt features. Thirdly, our approach emphasises data analysis and interpretation, focusing on the distinct pathological profiles and progression patterns in PD patients. Lastly, we incorporate advanced imaging techniques, such as super-resolved diffusion-weighted MRI, to enhance the model's precision, facilitating an in-depth examination of brain structures involved in PD.

Furthermore, we will integrate Responsible Artificial Intelligence (RAI) principles to enhance the framework's reliability, transparency, and interpretability. Key to this is the use of scoop-based and methodology-based explainers, as highlighted in works by F. K. Došilović et al. (2018), E. Tjoa et al. (2021), and Ali. S et al. (2023). These tools are essential for demystifying complex machine learning models. Scoop-based explainers, such as LIME (Ribeiro, M. et al., 2016) and SHAP (Lundberg, S. et al., 2017), provide valuable insights. LIME offers local understanding by showing how individual features affect predictions in specific cases, while SHAP gives a global perspective, highlighting the influence of features across the dataset. Methodology-based explainers complement these by delving deeper into the model's internal mechanisms. Backpropagation-based methods (Das, A. et al., 2020) and perturbation-based methods (Ivanovs, M. et al., 2021) examine the impact of various inputs on the outputs, an essential aspect for medical research. The integration of these explainers into our framework, combined with continuous feedback from researchers, ensures that our model meets the specific needs of medical research. It remains dynamically adaptive, addressing the 'black box' challenge inherent in advanced machine learning models.

Alternative fallback solutions

(i) Bayesian probabilistic graphical models in Python would enable us to longitudinally model the complex relationships between clinical and MRI biomarkers in different regions of interest, and graphically represent the structural and functional relationships between them. Bayesian probabilistic graphical models will also enable the creation of individualized models (digital twins) that reflect patients' unique pathological profiles. This is particularly relevant in the study of heterogeneity, as it enables personalized modelling and understanding of individual progression patterns.

(ii) Following on from some of the methods in the article by Di Folco et al. (2023), but this time with a central focus on imaging data, it would be interesting to create a multimodal map of PD evolution according to DWI-MRI markers (mean, radial and axial diffusivities, apparent diffusion coefficient, fractional anisotropy, free water) according to Parkinsonian patients’ age in our main regions of interest: dorsal motor nucleus, raphe nucleus, locus coeruleus, pedunculopontine nucleus, substantia nigra, BNM, diagonal band nucleus, amygdala, entorhinal cortex and piriform cortex.

Segmentation of regions of interest necessary for the alternative solutions: Segmentation of brainstem and basal forebrain nuclei affected in PD will be performed using multi-contrast templates per subject condition (with ANTs) using the super-resolved diffusion, neuromelanin-sensitive, T1-weighted, and QSM images to optimize 3D image quality and segmentation of small regions of interest, including structures related to the olfactory circuit and brainstem nuclei. An automated deep learning based multi-contrast approach will be developed. Using super-resolved DW images, we will perform tractography (building bundles between brain structures and their evolution using fractional anisotropy on super-resolved DW-MRI image templates) between structures related to sleep, olfactory and cortical circuits to study abnormalities and test our hypotheses.

Preliminary data:

“Super-resolution” DTI maps obtained from IQT-RF.

Typical PD progression of 4 imaging biomarkers and 8 clinical markers in the multimodal course map (Di Folco et al., 2023).

Assignments:

With the help of Romain Valabregue and Stéphane Lehericy, the recruited person will be working with the Center for NeuroImaging Research (CENIR). This study is related to a new and extremely ambitious collaborative project between {Aramis - Inria} and the teams Movement, Investigations and Therapeutics (Mov'it), and the “Molecular Pathophysiology of Parkinson's Disease” of ICM.

Responsibilities:

The person recruited is responsible for the representation learning methodology and the NEURACLE system and will take initiatives for including specific and targeted research instances as Parkinson’s Disease (first major instance) and other neurodegenerative diseases.

Feasibility/risk balance:

Neuracle is designed to become a game-changer in neuroimaging. Our teams have access to an important brain MRI dataset of about 315 patients (ICEBERG), PPMI dataset includes 1425 subjects and UkBioBank offers additional access to about 153 569 patients. All together, these data would consolidate a unique MRI dataset, allowing us to elaborate a novel approach, based on MRI representation learning. Unsupervised approaches will allow us to build an unbiased, generic semantic indexing, which will constitute a reference for a wide scientific community, a new way of dealing with this essential biomarker. The representation learning algorithms exist (https://github.com/facebookresearch/dinov2), in a community tested and validated version (Apache License Version 2.0). The risk is more at the data side, particularly the balance of the data categories (especially for under-represented categories). Acquisition heterogeneity devices are also to be considered as a risk factor. Concerning this aspect, the project teams (Botani. S, 2022) have a consistent experience in data quality control which could be used to overcome this obstacle. As a final remark, we consider that this project presents a configuration of average-risk coupled with an extremely high-gain profile.

Technical resources:

Integration of high-end computational resources from ICM, INRIA-CLEPS, and Jean-Zay supercomputers: specifically, the Nvidia GPU nodes A100 80GB and V100 32GB-16GB configurations.

Technical biases:

Self-supervised learning in representation learning is affected by biases in training data. To address this, a comprehensive review of training datasets is planned to detect biases like class imbalance. Strategies such as modifying the loss function in gradient backpropagation are used to give more weight to underrepresented conditions. Additionally, the model provides confidence scores for its predictions, enhancing its trustworthiness. In cases of new, unseen conditions, the model issues a cautionary disclaimer. These steps improve the model's robustness, reliability, transparency, and user-friendliness.

Human expertise:

Our team is skilled in using resources like ICM, INRIA-CLEPS, and Jean-Zay supercomputers, handling large datasets in multi-GPU environments. Proficient in PyTorch, CUDA, and CuDNN, we optimise tasks on Nvidia A100 and V100 GPUs. Our expertise in PyTorch enables us to develop efficient machine learning models, maximising GPU parallel processing for high efficiency in data handling and algorithm execution, effectively addressing large-scale computational challenges.

Ethical approvals:

We have the necessary ethical authorisations to continue to carry out MRI (3-Tesla imager) and clinical (ICM clinical investigation centre) examinations for the ongoing ICEBERG cohort (ending in 2025) on healthy, parkinsonian and iRBD participants, as well as freely using ICEBERG, PPMI and UkBioBank cohorts’ data.

Expected results:

This framework will assist researchers in building hypotheses by identifying differences and/or similarities in specific regions of interest through heatmaps. Furthermore, it will foster a collaborative dialogue between experts and the model. Indeed, merging experts’ hypotheses with the model's findings will provide means for refining hypotheses and expanding medical knowledge.

The expected results can be structured around the following elements:

• Unbiased Representation Learning Method and Models - the protocol and the model used to generate the virtual model of the brain (brain oracle)
• Unsupervised-generated virtual brain, digital neuroimaging oracle - the core of the NEURACLE project: the reference digital oracle.
• PD classification, prediction - PD classification and prediction from the NEURACLE
• PD propagation-route interpretation from NEURACLE
• Case-based reasoning - communication system between the expert and NEURACLE
• Retrieval system - retrieve similar cases from NEURACLE.
• Flexible representation-learning, generalisable to other neurological diseases - scalability and extension protocols (generate new questions and extend to a wider range of neurological diseases).

Steering/Management/Working arrangements:

The person recruited will be in charge of the operational design, validation and test of the representation learning framework, including the generation of a relevant feedback as to its integration into an active learning process. The work plan involves setting-up the dataset, before applying a dataset quality-check and correction (when necessary). Therefore, the first representation learning protocols will be studied for the MRI T1-weighted modality (the easier to access without any pre-processing or super-resolution). In a second phase, we will study the RL on other modalities: Neuromelanin-sensitive MRI, DW-MRI (after pre-processing and super-resolution steps) and QSM. Finally, after integrating the clinical information, we will proceed to a clinical test and validation phase, before expanding the capabilities of our model by including responsible AI features.

Potential impact and innovation:

By integrating clinical expertise with cutting-edge imaging techniques and artificial intelligence, this project promises several innovations with far-reaching implications. The exploration of PD progression, integrating clinical and MRI biomarkers, offers the possibility of tailoring interventions according to individual characteristics. This project contributes to innovation in neuroimaging using advanced techniques such as super-resolution, segmentation and tractography on DW-MRI. These techniques have the potential to improve diagnostic accuracy and provide more nuanced monitoring of disease progression. In the field of artificial intelligence, the integration of Representation Learning models represents a frontier in machine learning applications for clinical neuroscience and radiology. This could lead to more sophisticated predictive models and data-driven insights, potentially transforming our approach to neurological disorders. Much more than a digital twin, vision transformers represent a digital oracle enabling hypotheses to be tested from bulk data, unlike traditional engineering, which forces data to be structured according to project lineage. In addition, the integration of these techniques into database management introduces a new approach to disease modelling. This innovation facilitates the creation of virtual patient profiles, which could improve simulation studies and treatment planning.

Strategy to pursue the scientific program: 

If this method proves effective in distinguishing multicontrast MRI images and seeing differences in pathological brain activation, as well as providing a better understanding of propagation pathways, vision transformers could be applied to many other medical neuroimaging research cohorts, including Neuropreims (ICM/Hôpital Pitié-Salpêtrière) from 2025 onwards, involving 1,000 subjects over a 4-year period, and potentially become a reference tool in the reconstruction and synthesis of brain MRI data, and possibly DaTscan images to study the amount of dopamine in regions of interest.

This project will naturally be extended to other neurological diseases. Indeed, the powerful semantic indexing of a massive IRM dataset is without any precedence yet and will allow us to trigger new collaborations to extend its use to a wide variety of neurological pathologies. Besides, this is also likely to inspire new opportunities and will participate to increase the outreach of our institute.

References:

Alexander Selvikvåg Lundervold, Arvid Lundervold, (2019) An overview of deep learning in medical imaging focusing on MRI, Zeitschrift für Medizinische Physik, Volume 29, Issue 2, Pages 102-127,

Alexander et al. (2017) Image quality transfer and applications in diffusion MRI, Neuroimage; 152:283-298. Ali, S., Abuhmed, T., El-Sappagh, S., Muhammad, K., Alonso-Moral, J. M., Confalonieri, R., ... & Herrera, F. (2023). Explainable Artificial Intelligence (XAI): What we know and what is left to attain Trustworthy Artificial Intelligence. Information Fusion, 99, 101805.

Benoit Dufumier. Apprentissage des représentations en neuroimagerie : transfert de connaissance à partir de larges jeux de données contrôles vers de petites cohortes cliniques. Computer Vision and Pattern Recognition [cs.CV]. Université Paris-Saclay, 2022. English. ⟨NNT: 2022UPASG093⟩.

Bohnen, N., Postuma, R., B. (2020) Body-first versus brain-first biological subtyping of Parkinson’s disease, Brain, Volume 143, Issue 10, Pages 2871–2873

Simona Bottani. Machine learning for neuroimaging using a very large scale clinical datawarehouse. Artificial Intelligence [cs.AI]. Sorbonne Université, 2022. English. ⟨NNT: 2022SORUS110⟩.

Braak, H., Tredici, K. D., Rüb, U., de Vos, R. A. I., Jansen Steur, E. N. H., & Braak, E. (2003). Staging of brain pathology related to sporadic parkinson’s disease. Neurobiology of Aging, 24(2), 197–211. Borghammer P. (2018). How does parkinson’s disease begin? Perspectives on neuroanatomical pathways, prions, and histology. Mov Disord; 33:48–57.

Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., & Joulin, A. (2021). Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 9650-9660).

Couronné, R. Progression models for Parkinson’s Disease. (2021) Modeling and Simulation. PhD thesis, Sorbonne Université.

Darcet, T., Oquab, Maxime, MAIRAL, Julien, et al. (2023) Vision transformers need registers. arXiv preprint arXiv:2309.16588.

Das, A., & Rad, P. (2020). Opportunities and challenges in explainable artificial intelligence (xai): A survey. arXiv preprint arXiv:2006.11371.

Di Folco, C., Couronné, R., Arnulf, I., Mangone , G., Leu-Semenescu, M., Dodet, P., Vidailhet, M.; Corvol, J-C., Lehéricy, S. (2023) Charting disease trajectories from isolated REM sleep behavior disorder to Parkinson’s disease. Mov Disord. doi: 10.1002/mds.29662. Epub ahead of print. PMID: 38006282.

Došilović,F. K., Brčić, M. and Hlupić, N. (2018) "Explainable artificial intelligence: A survey," 2018 41st International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO), Opatija, Croatia, pp. 0210-0215, doi: 10.23919/MIPRO.2018.8400040.

Doty, R. L. (2012). Olfactory dysfunction in parkinson disease. Nature Reviews Neurology, 8(6), 329–339. García-Lorenzo D, Longo-Dos Santos C, Ewenczyk C, Leu-Semenescu S, Gallea C, Quattrocchi G, Pita Lobo P, Poupon C, Benali H, Arnulf I, Vidailhet M, Lehericy S. (2013) The coeruleus/subcoeruleus complex in rapid eye movement sleep behaviour disorders in REM sleep behaviour disorders in Parkinson’s disease. Brain Jul;136(Pt 7):2120-9.

Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., ... & Guo, B. (2021). Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision (pp. 10012-10022).

Gaurav, R. et al. (2022) ‘NigraNet: An Automatic Framework to assess nigral neuromelanin content in early parkinson’s disease using Convolutional Neural Network’, NeuroImage: Clinical, 36, p. 103250.

Horsager, J. et al. (2020) ‘Brain-first versus body-first parkinson’s disease: A multimodal imaging case-control study’, Brain, 143(10), pp. 3077–3088.

Ivanovs, M., Kadikis, R., & Ozols, K. (2021). Perturbation-based methods for explaining deep neural networks: A survey. Pattern Recognition Letters, 150, 228-234.

Kalia, L.V., Lang, A.E. (2015). Parkinson’s disease. The Lancet; 386:896–912.

Lundberg, S. M., & Lee, S. I. (2017). A unified approach to interpreting model predictions. Advances in neural information processing systems, 30.

Oquab, M. et al. (2023) ‘DINOv2: Learning Robust Visual Features without Supervision’, arXiv, pp. 1–31. doi: https://doi.org/10.48550/arXiv.2304.07193.

Parkkinen, L., Pirttila, T., Alafuzoff, I. (2008). Applicability of current staging/categorization of alpha-synuclein pathology and their clinical relevance. Acta Neuropathol; 115: 399-407

Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., ... & Sutskever, I. (2021, July). Learning transferable visual models from natural language supervision. In International conference on machine learning (pp. 8748-8763). PMLR.

Ribeiro, M. T., Singh, S., & Guestrin, C. (2016, August). " Why should i trust you?" Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining (pp. 1135-1144).

Tjoa, E. and Guan,C. (2021) "A Survey on Explainable Artificial Intelligence (XAI): Toward Medical XAI," in IEEE Transactions on Neural Networks and Learning Systems, vol. 32, no. 11, pp. 4793-4813

Ubeda-Bañon, I. et al. (2014) ‘Α-synuclein in the olfactory system in parkinson’s disease: Role of neural connections on spreading pathology’, Brain Structure and Function. Sep;219(5):1513-26.

Uchihara T., Giasson, B.I. (2016). Propagation of alpha-synuclein pathology: hypotheses, discoveries, and yet unresolved questions from experimental and human brain studies. Acta Neuropathol; 131: 49-73

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.

Main Activities (up to 5):

1. Develop reliable solutions for optimizing 3D visual transformers tailored to handling massive volumetric MRI datasets.
2. Design and prototype an experimental platform dedicated to neuroimage representation learning.
3. Conduct iterative testing and refinement of the neuroimaging representation learning platform until validation.
4. Develop and prototype programs, applications, and interfaces for the NEURACLE interrogation system, focusing on investigating Parkinson's Disease propagation pathways.
5. Conduct iterative testing and refinement of the NEURACLE interrogation system for studying Parkinson's Disease propagation pathways until validation.

Additional Activities (up to 3):

1. Adapt the initial prototype programs, applications, and interfaces of the NEURACLE interrogation system to enhance flexibility for addressing other neurodegenerative challenges.
2. Conduct iterative testing and refinement of the NEURACLE interrogation system to ensure its effectiveness in addressing various neurodegenerative challenges.
3. Prepare and submit publications, intellectual property documentation (if applicable), and reports documenting the project's findings and advancements.

Technical Skills and Proficiency Level Required:
- Proficient hands-on experience in deep learning and Python programming, including expertise with frameworks such as PyTorch.
- Advanced knowledge in Machine Learning and Pattern Recognition.
- Expertise in BioMedical Image Analysis.
- Strong foundation in applied mathematics and computer science.
- Familiarity with biomedical imaging techniques.

Language Proficiency:
- English: Expert level proficiency in both oral and written communication.
- French: Good proficiency level suitable for both oral and written communication.

Relational Skills:
- Strong oral communication skills for scientific discourse.
- Demonstrated ability in written communication through publications.

Other Valued Attributes:
- Proven ability to explore and devise solutions to complex problems.
- Collaborative mindset with a strong inclination towards teamwork.
- Positive and constructive attitude towards challenges and collaborations.
- Curiosity and eagerness to delve into medical applications.
- Willingness to learn more about Parkinson's and neurodegenerative diseases.

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking
• Flexible organization of working hours (after 12 months of employment)
• Professional equipment available (videoconferencing, loan of computer equipment, etc.)
• Social, cultural and sports events and activities
• Access to vocational training
• Social security coverage

According to civil service salary scales