IVIrene E. Vignon-Clementel

Irene E. Vignon-Clementel

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Irene E. Vignon-Clementel has a total of 268 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Hemodynamic insights into TIPS intervention for portal hypertension management: a comprehensive computational study

    by Pavlos Varsos, Friederike Schäfer, Cristina Ripoll, Nicolas Golse, Irene E. Vignon-Clementel

  • Voronoi-Based Myocardial Perfusion Region Prediction Enhanced by a Graph Convolutional Network

    by Mohamed Ali Srir, Raoul Sallé de Chou, Sabrina Lynch, Hugues Talbot, Laurent Najman, Matthew Sinclair, Irene Vignon-Clementel

  • Open Access

    Improved Physics-informed neural networks loss function regularization with a variance-based term

    by John M. Hanna, Hugues Talbot, Irene E. Vignon-Clementel

    In machine learning and statistical modeling, the mean square or absolute error is commonly adopted as an error metric, also called a "loss function". Although effective in reducing average error, this approach may fail to address localized outliers, leading to significant inaccuracies in regions with sharp gradients or discontinuities. This issue is particularly evident in physics-informed neural networks (PINNs), where such localized errors are expected and affect the overall solution. To overcome this limitation, we propose a novel loss function that combines the mean and standard deviation of the chosen error metric. By minimizing this combined loss function, the method ensures a more uniform error distribution and reduces the impact of localized high-error regions. The proposed loss function is easy to implement and tested in problems of varying complexity: 1D Poisson equation, unsteady Burger’s equation, 2D linear elastic solid mechanics, and 2D steady Navier Stokes equations. The results demonstrate improved solution quality and lower maximum error compared to the standard mean-based loss, with minimal impact on computational time.

    • 19 December 2025
    • Published by HAL in HAL
    Open Access

    Multi-fidelity graph-based neural networks architectures to learn Navier-Stokes solutions on non-parametrized 2D domains

    by Francesco Songia, Raoul Sallé De Chou, Hugues Talbot, Irene E Vignon-Clementel

    We propose a graph-based, multi-fidelity learning framework for the prediction of stationary Navier–Stokessolutions in non-parametrized two-dimensional geometries. The method is designed to guide the learning pro-cess through successive approximations, starting from reduced-order and full Stokes models, and progressivelyapproaching the Navier–Stokes solution. To effectively capture both local and long-range dependencies in thevelocity and pressure fields, we combine graph neural networks with Transformer and Mamba architectures.While Transformers achieve the highest accuracy, we show that Mamba can be successfully adapted to graph-structured data through an unsupervised node-ordering strategy. The Mamba approach significantly reducescomputational cost while maintaining performance.Physical knowledge is embedded directly into the architecture through an encoding - processing - physics in-formed decoding pipeline. Derivatives are computed through algebraic operators constructed via the WeightedLeast Squares method. The flexibility of these operators allows us not only to make the output obey the gov-erning equations, but also to constrain selected hidden features to satisfy mass conservation. We introduceadditional physical biases through an enriched graph convolution with the same differential operators describ-ing the PDEs. Overall, we successfully guide the learning process by physical knowledge and fluid dynamicsinsights, leading to more regular and accurate predictions

  • Derivation and Validation of Compartment Models: Implications for Dynamic Imaging

    by Jérôme Kowalski, Lorenzo Sala, Dirk Drasdo, Irene Vignon-Clementel

    Compartment models are mathematical representations of the transport of a chemical substance in the human body, assuming uniform concentration within each compartment, which corresponds to a distinct body part. Widely used in pharmaceutical and medical imaging through pharmacokinetic and tracer kinetic (TK) models, they are formulated as systems of time-dependent ordinary differential equations. However, these models do not account for the spatial dependence of physiological processes within individual compartments. This work aims to explore the physical interpretation of these models through mathematical derivation and to analyze their limitations. Three TK models are derived from more complex models regarding the physiological processes represented. The derivation introduces the hypotheses relevant to these processes. The most notable is the well-mixed hypothesis, resembling concentration homogeneity in the region of interest. The hypotheses are numerically tested by simulating the more complex model and evaluating the reduced models’ residuals. For a given tracer molecule, the validity of TK models is found to decrease as voxel size increases. This work proposes an algorithm to compute the maximum voxel size, above which the reduced models’ residuals exceed a chosen error threshold. For a 3 % error, voxel sizes larger than 250 μm are found critical for highly permeable vessels, while sizes smaller than 4 mm induce less than 3 % error in non-permeable vessels. Tested on literature-derived datasets, these findings identify diffusion as the underlying phenomenon leading to well-mixed compartments, and highlight the need for high spatial resolution imaging to improve TK parameters ground truth estimation.

  • A comparative study of lumped heart models for personalized medicine through sensitivity and identifiability analysis

    by Marie Haghebaert, Pavlos Varsos, Roel Meiburg, Irene Vignon‐Clementel

    Numerical modelling of the cardiovascular system is becoming an increasingly accepted tool for clinical applications, with the ultimate goal of personalized medicine. Lumped parameter models are attractive due to their low computational cost but often do not directly incorporate physical properties, thus requiring calibration to (often sparse) clinical data. Furthermore there exists a trade‐off between physiological relevance and model complexity, making the choice of cardiac model non‐trivial. For two established cardiac chamber models embedded in a haemodynamics whole circulation model, we perform sensitivity and identifiability analyses to examine the possibility of finding a unique subset of important parameters with varying levels of clinically measurable data, thereby examining their applicability in personalized medicine. To provide a concrete clinical context, the case of treatment planning for a young pulmonary arterial hypertension patient is considered. The methodology is however relevant for other pathophysiologies. The results suggest that the single‐fibre model, although a priori more complex than the time‐varying elastance model, is more amenable for patient‐specific modelling. This was found for representing the patient state from clinical data, defining parameter ranges for sensitivity analysis (SA), and in the results of the identifiability analysis. SA also revealed the most influential parameters, which for the right (respectively left) heart chambers, mostly affect the haemodynamics of these chambers and the pulmonary (respectively systemic) circulation but also the ones of the left (respectively right) side. This highlights the importance of studying the whole circulation, especially in diseases traditionally thought to affect only one side. image Key points: Two established cardiac chamber computational models are compared in the context of pulmonary hypertension, but the methodology holds for other pathophysiologies. Comprehensive patient‐specific data, accurate parameter ranges and thorough model validation are essential to enhance parameter identifiability, improve model personalization and eventually lead to better prediction of haemodynamic changes after clinical intervention. Due to its inherent mathematical constraints and physiologically interpretable input parameters, the single‐fibre model is easier to make patient‐specific and more accurately captures the non‐linear dynamics of ventricular pressure–volume loops than the simpler time‐varying elastance model, making it more suitable for personalized cardiac simulations. Accurate representation of both the left and right heart chambers in cardiac modelling is important, as our results suggest that parameters of each side impact the haemodynamics of the other circulatory side, making it worthwhile to measure the characteristics of both the right and left chambers, even if the disease emerges from one side.


  • Open Access

    Automatic Generation of Liver Virtual Models with Artificial Intelligence

    by Omar Ali, Alexandre Bône, Caterina Accardo, Belkacem Acidi, Amaury Facque, Paul Valleur, Chady Salloum, Marc-Michel Rohe, Irene Vignon-Clementel, Eric Vibert, et al.

    Objective: The clinical aim of this work is to predict intraoperative LRC from preoperative CT scans only. Summary of Background Data: Liver resection (LR) is the most prevalent curative treatment for primary liver cancer, yet overall mortality/morbidity rates remain elevated. The conventional definition and classification of LR complexity (LRC) lack inclusion of the disease-induced 3D anatomical surgery complexity. Methods: 3D models of the organ, tumors and blood vessels were generated from Deep Learning models trained on patients CT scans. The surgeons’ expertise on which anatomical factors lead to LRC was translated into a new anatomical frame of reference around the Hepatic Central Zone (HCZ). A fully automatic pipeline to generate the HCZ and quantify the tumors position relative to it was assessed. An AI model was then trained to predict LRC from a patient cohort for whom LRC was annotated at the end of each surgery. The AI-prediction was finally compared to prediction of surgeons that only saw the patient preoperative CT scan. Results: The 3D reconstructions are successfully evaluated on benchmark datasets. The HCZ is accurately generated for a variety of atypical vascular anatomies (dice score 82±4.6%). The automatic pipeline is successfully run on a 145 HCC patient cohort. The predicted LRC outperforms the surgeons’ individual and combined anticipated complexities (accuracy and AUC scores: 79.4±3.4% and 85.1±3.2% respectively). Conclusion: This automatic digital tool accurately predicts intraoperative LRC and paves the way for an innovative oncology surgery planning. This tool could help orient patients towards appropriate medical centers depending on the predicted LRC level.


    • 15 October 2024
    • Published by HAL in HAL
    Open Access

    Deep-learning and hemodynamics models: efficient complementarity for disease diagnosis and surgery planning

    by Irene Vignon-Clementel

    Est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d'enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.