I'm a machine learning and artificial intelligence research engineer specialized in deep learning and computer vision models applied to medical imaging. Currently working on a federated medical imaging research project at Henri Becquerel Cancer Center•AIMS Lab.
Specialized in Computer VisionMedical ImagingGenerative ModelingPET and MRI Imaging
Developing end-to-end architectures to monitor patient state and disease evolution by fusing Deep Learning, Generative AI, and Multi-Omics representations
Unify every modality
MRI, PET, CT
High-throughput
EHR, History
DNA, Proteomics
Lifestyle, Context
Generative & predictive models
From synthesis to forecasting
MRI to PET
Lesion delineation
Site normalization
Evolution prediction
Building robust deep learning models applied to clinical oncology and industrial computer vision
Research engineer on the PRT-K Federated-PET project, focused on developing federated multicenter solutions for PET image harmonization.
Design of robust generative augmentation models (VAE, Diffusion) to spatially and longitudinally model brain cancer (gliomas) evolution from rare and heterogeneous clinical acquisitions.
Design of semantic architectures and facial recognition models applied to the automotive sector and global HR systems post-merger.
A rigorous academic path with continuous honors in computer science, computer vision, and AI
Thesis: 'Generative models for predicting cancer progression from multimodal data'. Defended with praise on Sept 16, 2025.
High specialization in deep learning, NLP, and statistical algorithms. Graduating Valedictorian (Ranked 1st) with ~15/20.
Comprehensive study of distributed AI systems and embedded intelligence. Deep dive into hardware acceleration (Arduino, FPGA, SIMD vectorization) and advanced optimization techniques. Graduating Valedictorian (Ranked 1st) with ~15/20.
Acquired software development paradigms, design patterns, and completed a thesis using Deep Belief Networks (DBN) for NLP-based road accident tracking.
Selected peer-reviewed conference, journal and preprint papers
A. Kebaili, T. Carlier, A. Girard, A. Devillers, O. Humbert, S. Hapdey, C. Boucherie, F. Orlhac, P. Decazes
Aim/Introduction: EARL harmonization from raw PET data is critical for multicenter quantitative [18F]FDG PET/CT. We investigated whether non-harmonized standard PET images could be retrospectively standardized into EARL-compliant reconstructions through a data-efficient deep learning framework. Materials and Methods: We developed a U-Net-based residual image-to-image translation framework (PseudoEARL-Net) that predicts a voxel-wise Δ-map between standard PET and EARL-reconstructed images. Conclusion: Our PseudoEARL-Net model enables robust high-fidelity pseudo-EARL translation from routine [18F]FDG PET/CT scans, supporting radiomic reproducibility and the inclusion of historical non-compliant datasets in multicenter studies.
A. Kebaili, R. Modzelewski, J. Lapuyade-Lahorgue, M. Fontanilles, S. Thureau, S. Ruan
We propose a multi-task diffusion architecture dedicated to the spatio-temporal prediction of glioma progression. This model jointly generates future FLAIR sequences and probabilistic evolution maps based on Signed Distance Fields (SDF), integrating uncertainty quantification. Evaluated on both public datasets and a private internal cohort, the approach confirms strong inter-center robustness.
View PublicationA. Kebaili, J. Lapuyade-Lahorgue, P. Vera, S. Ruan
We propose AMM-Diff, a solution for missing MRI modality imputation from any available combination using an Image-Frequency Fusion Network (IFFN). The method adopts an adaptive reconstruction strategy to produce complete and coherent modalities, providing robust segmentation performance even under incomplete acquisitions.
View PublicationA. Kebaili, J. Lapuyade-Lahorgue, P. Vera, S. Ruan
This paper extends latent diffusion models to multi-modal frameworks, simultaneously generating multiple MRI sequences and associated masks while allowing condition-based tumor characteristics. The model produces coherent multi-modal datasets that significantly enhance downstream multi-sequence segmentation tasks on BRATS.
View PublicationA. Kebaili, J. Lapuyade-Lahorgue, P. Vera, S. Ruan
We present an HVAE with discriminative regularization designed to focus generation on relevant regions of interest. By integrating Hamiltonian dynamics for better latent space exploration, the model generates artifact-free image/mask pairs that improve Dice scores in data-scarce segmentation regimes.
View PublicationA. Kebaili, J. Lapuyade-Lahorgue, S. Ruan
This review critically synthesizes and compares major generative model families (VAEs, GANs, diffusion models) applied to medical imaging data augmentation. We highlight trade-offs between visual quality, sample diversity, and computational cost, proposing practical clinical recommendations.
View PublicationLet's discuss computer vision challenges, multi-omics clinical research, or open ML research collaborations
If you are a clinician looking for AI assistance, a machine learning researcher, or a student looking for clinical imaging internships, feel free to drop a message.
I am continuously open to opportunities for clinical research collaborations, advanced engineering, or open-source projects. If you are looking for a passionate research engineer to design robust, ethical, and clinically applicable pipelines, feel free to reach out.