Postdoc M/F – Automated Detection of Epilepsy-Related Anomalies in Brain FDG PET Using Pseudo-Healthy Reconstruction

Job Description

Missions

The postdoctoral fellow will apply, validate and evaluate a set of complementary anomaly detection methods for brain FDG PET, with a specific focus on the detection of epilepsy-related hypometabolic areas in the context of drug-resistant epilepsy.

Activity

Scientific Background and Activities
PET Imaging in Drug-Resistant Epilepsy
FDG PET has become a key modality in the pre-surgical workup of drug-resistant epilepsy, as it can reveal areas of cerebral hypometabolism associated with epileptogenic zones. In clinical practice, however, PET images are analysed visually, and the sensitivity and specificity of this approach greatly depend on the observer's experience. Abnormalities in epilepsy patients are often subtle and spatially diffuse, making them particularly difficult to identify reliably. There is therefore a strong clinical need for automated, data-driven tools capable of localising these anomalies objectively and supporting surgical planning.

Pseudo-Healthy Reconstruction for Anomaly Detection
Our team has been developing a family of anomaly detection methods based on the construction of a subject-specific pseudo-healthy image: at inference, the patient's image is compared to its pseudo-healthy reconstruction, producing an abnormality map that highlights areas deviating from what would be expected in the absence of pathology. In the context of drug-resistant epilepsy, such maps could help clinicians better localise epileptogenic zones and ultimately improve patient outcomes.

Registration and fusion approaches. Traditional voxelwise approaches register the subject's PET to a standard space and compare it to a population of controls using z-scores. However, their sensitivity is limited by inter-subject variability in non-pathological tracer uptake. To address this, subject-specific approaches were proposed that register multiple healthy controls to the target subject and combine them through image fusion to create a healthy appearance model [1]. While these methods have become less common due to concerns about speed and accuracy, recent advances in deep learning-based registration have improved both aspects, making it worthwhile to revisit them. Recent work in our group has demonstrated the potential of this modernised registration-based strategy for pseudo-healthy FDG PET synthesis [2].

Deep generative model approaches. More recent anomaly detection approaches use deep generative models capable of reconstructing pseudo-healthy images directly. A benchmark of VAE variants for pseudo-healthy reconstruction of 3D brain FDG PET established that the approach is feasible even with small training sets [3]. Further work has addressed key methodological challenges, including the impact of model variability on detection reliability [4], a fundamental trade-off between reconstruction quality and anomaly detection performance [5], and inference-time optimisation strategies to produce reconstructions that are both healthy-looking and subject-specific [6].

Objectives of the Postdoc
Building on this body of work, the postdoctoral fellow will apply, evaluate and compare these complementary anomaly detection approaches on FDG PET data from patients with drug-resistant epilepsy, in close collaboration with clinicians at the Pitié-Salpêtrière hospital. The abnormality maps generated will be evaluated for their ability to localise epileptogenic zones and support pre-surgical planning, using both public datasets and locally managed clinical data.

References
[1] Burgos, N., Cardoso, M. J., Samper-González, J., Habert, M.-O., Durrleman, S., Ourselin, S., & Colliot, O. (2021). Anomaly detection for the individual analysis of brain PET images. Journal of Medical Imaging, 8(2), 024003. https://doi.org/10.1117/1.JMI.8.2.024003
[2] Krause, M., Roy, H., & Burgos, N. (2026). Pseudo-healthy FDG PET synthesis: Revisiting a registration and fusion approach with deep learning. OHBM 2026. https://inria.hal.science/hal-05698660
[3] Hassanaly, R., Solal, M., Colliot, O., & Burgos, N. (2025). Benchmarking 3D generative autoencoders for pseudo-healthy reconstruction of brain 18F-fluorodeoxyglucose positron emission tomography. Journal of Medical Imaging, 12(5), 054005. https://doi.org/10.1117/1.JMI.12.5.054005
[4] Solal, M., André, P., & Burgos, N. (2026). Unsupervised anomaly detection in brain FDG PET with deep generative models: An experimental analysis of model variability and mitigation strategies. Medical Imaging 2026: Image Processing, 13925, 32-50. https://doi.org/10.1117/12.3083687
[5] Senellart, A., Solal, M., Allassonnière, S., & Burgos, N. (2026). Mitigating the reconstruction-detection trade-off in VAE-based unsupervised anomaly detection. 2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI), 1-5. https://doi.org/10.1109/ISBI61048.2026.11516028
[6] Solal, M., Senellart, A., Burgos, N., & Allassonnière, S. (2026). Latent maximum-a-posteriori approach to improve pseudo-healthy reconstruction quality. OHBM 2026. https://hal.science/hal-05560394

Your Profil

Skills

- PhD in medical image analysis, computer science, applied mathematics, or a related field
- Expertise in anomaly detection using deep generative models (in particular VAEs) and/or image registration
- Programming skills in Python and PyTorch
- Experience with 3D brain neuroimaging data, in particular FDG PET
- Good communication skills and ability to work in an interdisciplinary team
- Proficiency in English required

Your Work Environment

You will work within the ARAMIS team (www.aramislab.fr) at the Paris Brain Institute (https://institutducerveau-icm.org), one of the world's leading research institutes for neurosciences. The institute is ideally located at the heart of the Pitié-Salpêtrière hospital, in the centre of Paris. The ARAMIS team, which is also part of Inria (the French National Institute for Research in Digital Science and Technology), is dedicated to the development of new approaches for the analysis of large neuroimaging and clinical datasets. You will interact locally with the PhD students, postdoctoral fellows and engineers of the team, as well as with our collaborators at the Pitié-Salpêtrière hospital. You will also interact with researchers of PR[AI]RIE, the Interdisciplinary Institute of Artificial Intelligence of Paris (https://prairie-institute.fr). The position will be supervised by Ninon Burgos, CNRS research director and co-head of the ARAMIS team.

Compensation and benefits

Compensation

Between 3,300€ and 4,200€ monthly gross, depending on experience

Annual leave and RTT

44 jours

Remote Working practice and compensation

Pratique et indemnisation du TT

Transport

Prise en charge à 75% du coût et forfait mobilité durable jusqu’à 300€

About the offer

About the CNRS

The CNRS is a major player in fundamental research on a global scale. The CNRS is the only French organization active in all scientific fields. Its unique position as a multi-specialist allows it to bring together different disciplines to address the most important challenges of the contemporary world, in connection with the actors of change.

CNRS

The research professions