Youness Mellak

LaTIM · Université de Bretagne Occidentale

I am a postdoctoral researcher at LaTIM, Université de Bretagne Occidentale in Brest, France. My work lies at the intersection of medical imaging physics, inverse problems, and machine learning.

I am particularly interested in scalable generative priors for three-dimensional reconstruction, direct PET reconstruction from detector measurements, positron-range correction, and physics-informed learning. My current research explores diffusion and flow models for PET, CT, and MRI at clinical volume sizes.

I completed my Ph.D. at LaTIM in 2025 with the thesis Deep learning approaches for PET imaging: three-gamma PET, positron range correction and direct reconstruction.

Code and material related to my work can be found on the publications page and on GitHub. The best way to reach me is by email.

experience

Postdoctoral ResearcherOct 2025 - present
LaTIM, INSERM UMR 1101 · Brest, France

Generative modeling and learned priors for large-scale 3D inverse problems, including low-dose PET, sparse-view CT, and undersampled MRI.

Ph.D. ResearcherOct 2021 - Sep 2025
LaTIM · Université de Bretagne Occidentale

Direct PET reconstruction, positron-range correction, Monte Carlo surrogate modeling, and direct three-gamma PET reconstruction.

Research InternMar 2021 - Sep 2021
Inria Sophia Antipolis · i3S · Nice, France

Learning-based analysis of volumetric OCT data, including severity grading, explainability, and three-dimensional instance segmentation.

news

Jul 18, 2026 3D Latent Implicit Diffusion for Scalable Volume Generation was accepted as a Lightning Talk and Poster at the 2026 IEEE NSS/MIC/RTSD conference.
Jul 18, 2026 Frozen Stable Diffusion VAE as a Slice-Wise Encoder for Scalable 3-D Medical Image Generation was accepted for a Poster presentation at the 2026 IEEE NSS/MIC/RTSD conference.

selected publications

  1. Continuous 3-D Latent Diffusion for Medical Generation and Reconstruction
    Youness Mellak, Antoine De Paepe, Dimitris Visvikis, and Alexandre Bousse
    Manuscript under review, 2026
    PDF Code Project
    Under review 2026

    Continuous 3-D Latent Diffusion for Medical Generation and Reconstruction

    • Youness Mellak1,
    • Antoine De Paepe1,
    • Dimitris Visvikis1,
    • Alexandre Bousse1
    • 1LaTIM, INSERM UMR 1101, Université de Bretagne Occidentale, Brest, France
    PDF Code

    Abstract

    High-resolution three-dimensional (3-D) medical diffusion models remain constrained by the cost of processing full volumes, even when denoising is performed in a compact latent space. We introduce a continuous 3-D latent diffusion model (LDM) framework for computed tomography (CT) and magnetic resonance imaging (MRI) generation and measurement-guided reconstruction. Its central component is a compact autoencoder (AE) with a coordinate-conditioned local implicit image function (LIIF) decoder that represents a volume as a continuous function of spatial coordinates. By evaluating the convolutional decoder once on the latent grid and restricting repeated computation to a lightweight implicit head, the proposed design avoids overlapping sub-volume decoding while remaining differentiable for inverse-problem optimization. We evaluate the framework on CT volumes of 512³ voxels and MRI volumes of 256³ voxels. On high-resolution CT, the proposed AE is approximately ×12–32 faster than the evaluated reference autoencoders, achieves the lowest peak graphics processing unit (GPU) memory use, and retains comparable structural fidelity despite a moderate reduction in voxel-level accuracy. The resulting frozen 3-D latent prior generates coherent full volumes without visible patch seams and can be applied, without task-specific retraining, to sparse-view CT and accelerated MRI reconstruction through hard data consistency. Although direct pixel-domain reconstruction remains more accurate, the results demonstrate that a single volumetric latent prior can support both unconditional generation and measurement-conditioned reconstruction on one GPU.

    Key results

    Central sagittal, coronal and axial slices of CT volumes generated by the proposed model, MAISI and 3D MedDiffusion, and of an unpaired CT volume
    Figure 1. Unconditional CT volume generation. Central sagittal, coronal and axial slices of samples from the proposed method, MAISI and 3D MedDiffusion, with an unpaired real CT volume for reference. All volumes use the same HU window.
    Sparse-view CT reconstructions from 30 noisy projections with FBP, DDS, the proposed method and the ground truth, with zoomed regions and error maps
    Figure 2. Sparse-view CT reconstruction from 30 noisy projections. For each method (FBP, DDS, ours, ground truth), the top panel shows the reconstruction and the region of interest; the bottom panels show its magnification and the absolute error map in HU.

    BibTeX

  2. Dual-Input Dynamic Convolution for Positron Range Correction in PET Image Reconstruction
    Youness Mellak, Alexandre Bousse, Thibaut Merlin, Élise Émond, Mikko Hakulinen, and Dimitris Visvikis
    IEEE Transactions on Radiation and Plasma Medical Sciences, 2025
    IEEE TRPMS 2025

    Dual-Input Dynamic Convolution for Positron Range Correction in PET Image Reconstruction

    • Youness Mellak1,
    • Alexandre Bousse1,
    • Thibaut Merlin1,
    • Élise Émond2,
    • Mikko Hakulinen3,
    • Dimitris Visvikis1
    • 1LaTIM, INSERM UMR 1101, Université de Bretagne Occidentale, Brest, France
    • 2Brainlab, Munich, Germany
    • 3Kuopio University Hospital, Kuopio, Finland

    Abstract

    Positron range (PR) blurs PET images, and the effect grows with higher-energy emitters such as ⁶⁸Ga and with the sub-millimetre resolution of recent scanners. In iterative reconstruction, PR can be modelled as an image-domain blurring operator B(μ), but this operator varies with the surrounding tissue and is usually approximated by a spatially invariant kernel, which underestimates the blur in low-density regions such as the lungs. Learned corrections that apply a CNN before each forward projection improve contrast but do not provide the matching backward operator. We model B(μ) as a transposed convolution with spatially varying kernels: a CNN predicts the PR point-spread function of each voxel from its local attenuation neighbourhood and a distance map. The network is trained on Monte Carlo (GATE) point-spread functions simulated in random lung, water and bone material patches. Because B(μ) and B(μ)ᵀ share the same predicted kernels, the operator plugs consistently into MLEM. On digital and XCAT phantoms, DDConv reproduces the Monte Carlo annihilation distributions at tissue interfaces where a spatially variant tissue-dependent (SVTD) model fails, and it was also evaluated in reconstructions of Monte Carlo and patient data.

    Method overview

    Diagram of the forward and transposed positron-range blurring operators sharing kernels predicted by the network G theta
    Figure 1. Positron-range blurring operators. Top: the transposed operator B(μ)ᵀ; bottom: the forward operator B(μ). Both use the spatially varying PSFs wⱼ predicted by the same model G_θ from the local attenuation map; the right side details the architecture of G_θ.

    Key results

    XCAT activity phantom with four lesions and its positron-range blurring by Monte Carlo, SVTD and DDConv
    Figure 2. Positron-range blurring of an XCAT ⁶⁸Ga activity phantom with four lesions: two in the lung, one at the lung/soft-tissue interface (3) and one at the lung/liver interface (4). SVTD fails to blur lesions 3 and 4 correctly, while DDConv stays close to the Monte Carlo reference.

    BibTeX

  3. Direct3γ: A Pipeline for Direct Three-Gamma PET Image Reconstruction
    Youness Mellak, Alexandre Bousse, Thibaut Merlin, Debora Giovagnoli, and Dimitris Visvikis
    IEEE Transactions on Radiation and Plasma Medical Sciences, 2025
    IEEE TRPMS 2025

    Direct3γ: A Pipeline for Direct Three-Gamma PET Image Reconstruction

    • Youness Mellak1,
    • Alexandre Bousse1,
    • Thibaut Merlin1,
    • Debora Giovagnoli1,
    • Dimitris Visvikis1
    • 1LaTIM, INSERM UMR 1101, Université de Bretagne Occidentale, Brest, France

    Abstract

    Three-gamma (3γ) PET aims to improve spatial resolution and reduce noise in nuclear medicine. The Direct3γ pipeline addresses the inherent challenges of 3γ PET systems, such as detector imperfections and uncertainty in photon interaction points. A key feature of the pipeline is its ability to determine the order of interactions through a model trained on Monte Carlo simulations using the GATE toolkit, which provides the information needed to construct Compton cones that intersect the line of response to estimate the emission point. The pipeline processes raw 3γ PET data, reconstructs histo-images by propagating energy and spatial uncertainties along the line of response, and applies a 3-D CNN to refine these intermediate images into high-quality reconstructions. To further enhance image quality, it combines supervised learning with adversarial losses, the latter preserving fine structural details. Direct3γ consistently outperforms conventional 200-ps TOF PET in terms of SSIM and PSNR.

    Method overview

    Direct3γ pipeline: raw three-gamma events, interaction ordering, histogrammer with attenuation map, histo-image, 3D U-Net, reconstructed image
    Figure 1. Direct3γ pipeline: (i) a modified interaction network (MIN) orders the photon interactions of each raw 3γ event; (ii) the 3γ histogrammer builds a histo-image using the attenuation map; (iii) a 3-D U-Net turns the histo-image into the reconstructed image.

    Key results

    Sagittal, coronal and transverse views of an XCAT phantom: ground truth, TOF, DIP-TOF and Direct3γ reconstructions
    Figure 2. Human-sized scanner, XCAT phantom: ground truth, 200-ps TOF PET, DIP-TOF and Direct3γ in sagittal, coronal and transverse views. The coronal ground truth shows the lesion ROIs (red) and background ROIs (green).

    BibTeX