Séminaires d'équipe

Organisation

Les activités d'équipes ont lieu le mardi à 11h en salle de réunion à Explora

Date des réunions permanents à 11h : not fixed

Calendar (Only invited have access to the calender)

Organisateur

Depuis le 2 février 2021 : Michel Duprez

Réservation des salles

Meeting room 114

Meeting room 127

Conference room at IHU

=Prochains séminaires / réunions

• Mardi 23 avril 2024 : Présentation de Nicola Zotto

Titre: "Boundary condition and shape independent prediction of soft tissue deformation"

Résumé: "Presentation of my work on shape and boundary condition independent features of hyper-elastic material deformations using Graph Neural Networks. We will first go over a quick introduction to the Graph Neural Networks architecture followed by the presentation of the various challenges we had to overcome in adapting this framework from "dynamic rollout" predictions to "static boundary problems."

• Vendredi 26 avril 2024 : Présentation de Duc Hoan NGUYEN

Titre: "Regularization in Reproducing Kernel Hilbert Spaces for Covariate Shift Domain Adaptation"

Résumé: "In machine learning, the domain adaptation problem arises when a core assumption of the statistical learning theory is violated. The above assumption supposes that previously seen and future inputs/outputs of the systems under consideration are governed by the same probability law. Since, in reality, it is not often the case, the domain adaptation problem becomes practically important. In the present study, we analyze the use of the general regularization scheme in the scenario of unsupervised domain adaptation under the so-called covariate shift assumption. Learning algorithms arising from the above scheme are generalizations of importance weighted regularized least squares method, which up to now is among the most used approaches in the covariate shift setting. We explore a link between the considered domain adaptation scenario and estimation of Radon-Nikodym derivatives in reproducing kernel Hilbert spaces, where the general regularization scheme can also be employed and is a generalization of the kernelized unconstrained least-squares importance fitting. We estimate the convergence rates of the corresponding learning algorithms and discuss how to resolve the issue with the tuning of their parameters. The theoretical results are supported by extensive numerical experiments with various datasets."

• Mardi 30 avril 2024 : Présentation de Long

Séminaires / réunions passés

• Mardi 16 avril 2024 : Présentation de Léo Bois

Titre: "Neural networks in numerical schemes for fluid simulations"

Résumé: "One of the challenges of fluid simulation is the occurrence of discontinuous solutions, which numerical schemes have to deal with. Various methods can be used to find a compromise between diffusion and oscillations at the discontinuities. In this presentation, I will mention two of them and explain how we have used neural networks to better tune these methods."

• Mardi 19 mars 2024 : Présentation de Cyril Meyer

Titre: "Segmentation d'images de microscopie électronique par apprentissage profond pour l'analyse quantitative de l'ultrastructure cellulaire."

Résumé: "Durant ce séminaire, je vous présenterai mon parcours, mes travaux de recherche et mes projets futurs. L'imagerie FIB-SEM permet d'obtenir des images 3D de résolution nanométrique en microscopie électronique cellulaire. Une segmentation de ces images est nécessaire pour permettre une analyse quantitative et comparative de l'ultrastructure cellulaire. Les méthodes actuelles basées sur les réseaux de neurones convolutifs rencontrent des défis liés à la rareté des données annotées, au bruit et aux variations d'acquisition. L'objectif de ce travail est de développer des méthodes de segmentation robustes. Deux axes principaux de recherche ont été explorés. Premièrement, l'amélioration de la qualité des segmentations en combinant apprentissage profond et morphologie mathématique, en particulier, les opérateurs connexes. Deuxièmement, la réduction du temps d'annotation nécessaire, en expérimentant des méthodes de segmentation interactive, mais aussi des méthodes automatiques entraînées à partir d'annotations faibles. En conclusion, je présenterai mon projet de recherche actuel, axé sur les deux pistes suivantes : la réduction de la quantité d'annotations requises lors de l'entraînement de modèles d’apprentissage profonds, et l'exploration de méthodes visant à rendre les modèles plus compacts et économes en ressources."

• Mardi 12 mars 2024 : Présentation de Camille Gontier (Université de Pittsburgh)

Titre: "Efficient sampling-based Bayesian Active Learning for synaptic characterization"

Résumé: "Bayesian Active Learning (BAL) is an efficient framework for learning the parameters of a model, in which input stimuli are selected to maximize the mutual information between the observations and the unknown parameters. However, the applicability of BAL to experiments is limited as it requires performing high-dimensional integrations and optimizations in real time. Current methods are either too time consuming, or only applicable to specific models. Here, we propose an Efficient Sampling-Based Bayesian Active Learning (ESB-BAL) framework [1], which is efficient enough to be used in real-time biological experiments. We apply our method to the problem of estimating the parameters of a chemical synapse from the postsynaptic responses to evoked presynaptic action potentials. Using synthetic data and synaptic whole-cell patch-clamp recordings, we show that our method can improve the precision of model-based inferences, thereby paving the way towards more systematic and efficient experimental designs in physiology. [1] Gontier, C., Surace, S. C., Delvendahl, I., Müller, M., & Pfister, J. P. (2023). Efficient sampling-based Bayesian Active Learning for synaptic characterization. PLoS computational biology, 19(8), e1011342. https://doi.org/10.1371/journal.pcbi.1011342"

• Jeudi 29 février 2024 à 11h : Présentation de Karim Makki, EnCoV-Institut Pascal, CNRS/Université Clermont-Auvergne

Titre: "La chirurgie assistée par ordinateur: de l'imagerie pré-opératoire, à la navigation 3D en temps réel."

Résumé: "En chirurgie assistée par ordinateur, l’image joue un rôle primordial. Les techniques d’imagerie médicale sont multiples et permettent aux cliniciens de bien suivre l'évolution d’une pathologie ainsi que la mise en oeuvre des bonnes pratiques tout au long du processus d’intervention chirurgicale: la phase préopératoire, la navigation en temps réel durant l'opération chirurgicale, ainsi que la phase postopératoire ou celle de suivi. Au-delà de l’acquisition de l’image, son exploitation nécessite de l’expertise supplémentaire tel que: le traitement des images dans le sens classique du terme, l’emploi des techniques de vision par ordinateur, ainsi que l’emploi des techniques modernes d’apprentissage profond. Un traitement pertinent conduit à la réalisation de certaines tâches telles que la segmentation de l’image, le recalage et la classification, etc… toujours utiles pour l’aide au diagnostic. Parmi les techniques d’imagerie sont: l’IRM dynamique qui permet de suivre le mouvement des articulations, des organes, et des tissus mous de façon non-invasive. Cette technique sert essentiellement à guider les cliniciens pendant les phases pré et post-opératoires. Ses principales limites sont: la basse résolution, ainsi que la présence quasi-permanente du bruit et des artefacts de mouvement dans les images temporelles. Tandis que l’endoscopie, consistant à pénétrer un long tube dans le corps du patient et une petite caméra sert à examiner l'intérieur du corps, permettant ainsi aux chirurgiens de voir ce qui se passe lors de l'opération en temps réel. Dans cette présentation, nous montrons l’utilité de quelques techniques de traitement d’image pour la reconstruction spatio-temporelle des séquences d’IRM dynamique pour la cheville (lors des mouvements dorsi-plantar flexion chez les enfants) et la vessie (lors d’une respiration profonde chez les femmes adultes). Par la suite, nous présentons la combinaison des techniques de vision par ordinateur classiques avec l’apprentissage profond pour la reconstruction 3D des organes observées par un endoscope à partir d’une séquence d’images 2D (pour le colon et le foie) pour une navigation plus riche et plus précise."

• Mardi 27 février 2024 à 11h : Présentation de Fethi Ghazouani

Titre: "Exploration de l'Apprentissage Profond pour la Fusion d'Images IRM Multimodales dans la Segmentation des Tumeurs Cérébrales"

Résumé: "Les transformers pour la segmentation des images médicales ont soulevé un grand intérêt. Contrairement aux réseaux convolutionnels (CNN), les transformers utilisent des auto-attentions qui n'ont pas un fort biais inductif. Cela donne aux transformers la capacité d'apprendre des dépendances à longue portée et des capacités de modélisation plus puissantes. Dans ce séminaire, j’exposerai mes travaux de recherches antérieures qui portent sur la fusion d'images IRM multimodales basée sur l’apprentissage profond multi-modal/tâche pour la segmentation et la prédiction de la tumeur cérébrale. Je présenterai en première partie l’architecture du modèle Swin Transformer et son principe de calcul de l’auto-attention, introduit pour réduire la complexité de l’opération d’auto-attention utilisée dans le modèle Vision Transformer (ViT) classique. Je détaillerai ensuite l’approche proposée pour la segmentation d’images multimodales de tumeurs cérébrales, qui s’appuie sur le modèle Swin Transformer et l’auto-attention locale améliorée (ELSA). Dans une seconde partie, je montrerai, comment il est possible de réaliser une fusion par corrélation multimodale dans l’espace des caractéristiques latentes pour la segmentation de la tumeur et prédiction de la localisation de la récidive de la tumeur cérébrale. Dans cette approche, le mécanisme d’apprentissage par transfert a été appliqué afin d’améliorer la performance et la précision du modèle dans le cas où les données sont limitées."

• Mardi 20 février 2024 : Présentation de Killian Vuillemot

Titre: "Phi-FEM-FNO: a new approach to train a Neural Operator as a fast PDE solver for variable geometries"

Résumé: "In this talk, we propose a way to solve partial differential equations (PDEs) by combining machine learning techniques and a new finite element method called phi-FEM. For that, we use the Fourier Neural Operator (FNO), a learning mapping operator. The purpose of this talk is to provide numerical evidence to show the effectiveness of this technique. We will focus here on the resolution of the Poisson equation with non-homogeneous Dirichlet boundary conditions. The key idea of our method is to address the challenging scenario of varying domains, where each problem is solved on a different geometry. Since we use the phi-FEM approach, we will consider domains defined by level-set functions. We will first recall the idea of phi-FEM and of the Fourier Neural Operator. Then, we will explain how to combine these two methods. We will finally illustrate the efficiency of this combination with some numerical results on two test cases."

• Mardi 30 janvier 2024 : Présentation de Kebing XUE

Titre: "Generation of human model in motion using spectral domain representation of 3D meshes."

Résumé: "In this work, we focus on human motion generation with the help of the diffusion model. Specifically, different from other works that use joint representation for training and require the regression of SMPL model to get shapes during sampling, we take advantage of the spectral representation of 3D meshes and train a diffusion model that is capable of generating the surfaces of the human model in motion directly. This will not only simplify the generation process but also allow us to control the generated shape explicitly and subtly."

• Mardi 23 janvier 2024 : Présentation de Amel Hidouri

Titre: "Symbolic Artificial Intelligence for Data Mining"

Résumé: "Data mining remains a vibrant field of research in Artificial Intelligence (AI), with active exploration focused on advancing techniques. Recently, it has found applications in both Explainable AI and the broader realm of machine learning. Pattern discovery, a well-established topic in data mining, holds diverse applications such as association rule mining, clustering, classification, and feature selection. When coupled with symbolic AI, it provides a declarative framework for discovering different kinds of patterns (e.g. frequent itemsets, sequence, gradual itemsets, etc). This presentation delves into multiple facets of pattern mining, encompassing tasks ranging from the computation of frequent and rare patterns to solving high utility itemset mining problems using symbolic AI. Following this exploration, we will introduce several notions related to explainability, covering both informal and formal approaches. The presentation concludes with a practical application in healthcare and industry."

• Mardi 16 janvier 2024 : Présentation de Paul Baksic

• Mardi 19 décembre 2023 : Présentation de Darshan Venkatrayappa

Titre: "Unidentified floating object detection in maritime environment"

Résumé: "In this work we present a new unsupervised approach to detect unidentified floating objects in the maritime environment. The proposed approach is capable of detecting floating objects online without any prior knowledge of their visual appearance, shape or location. Given an image from a video stream, we extract the self-similar and dissimilar components of the image using a visual dictionary. The dissimilar component consists of noise and structures (objects). The structures (objects) are then extracted using an a contrario model."

• Mardi 12 décembre 2023 : Présentation de Frédérique Lecourtier

Titre: "Development of hybrid finite element/neural network methods to help create digital surgical twins"

Résumé: "The aim of this work is to develop hybrid finite-element and neural-network methods for the creation of digital surgical twins. The idea is to train a neural network to predict the solution of a given problem, with the intention of injecting this solution into a finite element solver. The purpose of this work is to combine the speed of neural networks with the accuracy of finite element methods. Here, we will be using the PhiFEM immersed boundary method, which consists in immersing the geometry of an organ, defined by a Levelset function, in a regular mesh. The Finite Elements resolution step after the neural network prediction allows both correction and certification of the prediction."

• Mardi 5 décembre 2023 : Présentation de Vincent Italiano

Titre: "Augmented reality for minimally invasive liver surgery using digital twin"

Résumé: "The objective of this work is to provide surgeons with an enhanced view of the liver during minimally invasive operations. The project is divided in two parts: the pre-operative part, where a 3D model of the patient's liver is extracted from a CT scan to construct a digital twin using neural networks and FEM. Then the intra-operative part, which involves extracting real-time data from the liver during the operation to feed the numerical twin."

• Mardi 28 novembre 2023 : Présentation de Boyang Yu

Titre: "Inverse Garment and Pattern Modeling with a Differentiable Simulator"

Résumé: "The capability to generate simulation-ready garment models from 3D shapes of clothed humans will significantly enhance the interpretability of captured geometry of real garments, as well as their faithful reproduction in the virtual world. This will have notable impact on fields like shape capture in social VR, and virtual try-on in the fashion industry. To align with the garment modeling process standardized by the fashion industry as well as cloth simulation softwares, it is required to recover 2D patterns. This involves an inverse garment design problem, which is the focus of our work here: Starting with an arbitrary target garment geometry, our system estimates an animatable garment model by automatically adjusting its corresponding 2D template pattern, along with the material parameters of the physics-based simulation (PBS). Built upon a differentiable cloth simulator, the optimization process is directed towards minimizing the deviation of the simulated garment shape from the target geometry. Moreover, our produced patterns meet manufacturing requirements such as left-to-right-symmetry, making them suited for reverse garment fabrication. We validate our approach on examples of different garment types, and show that our method faithfully reproduces both the draped garment shape and the sewing pattern."

• Mardi 21 novembre 2023 : Présentation de Robin Enjabert

Titre: "Title : Physics-aware Human-object interaction"

Résumé: "As part of our collaborative project with ETRI (South Korea), the objective of this work is to be able to generate simulated movements of vitual human bodies performing everyday actions, including a physical cognition. This involves two main aspects : integrating 4D human data into a differentiable physics engine, and the implementation of a DL controller to generate realistic movements. Using an analytic policy gradients method, we use the differentiability of the simulator to train the DL controller and then generate a variety of daily activities."

• Mardi 14 novembre 2023 : Présentation de Valentina Scarponi

Titre: "FBG-driven simulation for predictive mixed reality during endovascular interventions"

Résumé: "Endovascular interventions are procedures designed to diagnose and treat vascular diseases, using catheters to navigate inside arteries and veins. Thanks to their minimal invasiveness, they offer many benefits, such as reduced pain and hospital stays, but also present many challenges for clinicians, as they require specialized training and heavy use of X-rays. This is particularly relevant when accessing (i.e., cannulating) small arteries with steep angles, such as most aortic branches. To address this difficulty, we propose a novel solution that enhances fluoroscopic 2D images in real-time by using virtual augmentation of the catheter and guidewire. In contrast to existing works, proposing either simulators or simple augmented reality frameworks, our approach involves a predictive simulation showing the resulting shape of the catheter after guidewire withdrawal without requiring the clinician to perform this task. Our system demonstrated accurate prediction with a mean 3D error of 2.4±1.3 mm and a mean error of 1.1±0.7 mm on the fluoroscopic image plane between the real catheter shape after guidewire withdrawal and the predicted shape. A user study reported an average intervention time reduction of 56% when adopting our system. The developed system can significantly reduce the time needed to access small vessels, thus lowering X-ray exposure and operating time. Our approach is well-suited for assisting with the effective cannulation of arteries branching off the aorta and has the potential to improve the efficacy, efficiency, and safety of almost all endovascular procedures."

• Mardi 24 octobre 2023 : Présentation de Diwei Wang

Titre: "Video-based 3D gait reconstruction for assessing Alzheimer’s Disease and Dementia with Lewy Bodies"

Résumé: "Dementia with Lewy Bodies (DLB) and Alzheimer’s Disease (AD) are two common neurodegenerative diseases among elderly people. Gait analysis plays a significant role in clinical assessments to discriminate these neurological disorders from healthy controls, to grade disease severity, and to further differentiate dementia subtypes. We propose a deep-learning based model specifically designed to evaluate gait impairment score for assessing the dementia severity using monocular gait videos. Specifically, we develop a gait reconstructor named MAX-GRNet to estimate the sequence of 3D body skeletons, and thereon, perform classification to determine the MDS-UPDRS gait scores. To design the MAX-GRNet, we initially explore the 3D pose corrections based on spatial-temporal gait features extracted from the input video, which helps to maintain the temporal coherence within walking sequences. Experimental results demonstrate that our technique outperforms alternative state-of-the-art methods. Nevertheless, the reconstruction is susceptible to loosing fine-grained details and over-smoothing irregular patterns. Subsequently, we leverage the mesh-aligned image features to optimize the precision of gait reconstruction."

• Mardi 10 octobre 2023 : Présentation de Pablo Alvarez

Titre: "Physically inspired regularization: application to deformable image registration"

Résumé: "Deformable image registration (DIR) of medical images has been a subject of research for decades, given its great potential for clinical applications such as interventional guidance and multi-modal image fusion, to name a few. DIR is an ill-posed problem, and various regularization strategies have been proposed to improve its convergence and resulting accuracy. However, these regularization approaches, although commonly used, derive from geometrical assumptions that may not be compatible with the deformation of organs. Consequently, physically inspired regularization schemes have been introduced. These methods are designed to mimic an organ's behavior as described by the laws of physics, and have achieved promising results. An important limitation, however, is their dependence on the finite-element method to compute kinematic quantities. I present a physically inspired regularization scheme that does not require the finite-element method, and can be easily integrated into classical deformable image registration algorithms. I illustrate the potential of such methodology using a SIREN neural network to register synthetic data from physical simulations, as well as breathing lung images from the publicly available DIRLab dataset. The implemented approach achieves target registration errors around 1.5 mm in average for this dataset, which is close to methods in the state-of-the-art."

• Mardi 3 octobre 2023 : Présentation de Sidaty El Hadramy

Titre: "Trackerless Volume Reconstruction from Intraoperative Ultrasound Images"

Résumé: "We propose a method for trackerless ultrasound volume reconstruction in the context of minimally invasive surgery. It is based on a Siamese architecture, including a recurrent neural network that leverages the ultrasound image features and the optical flow to estimate the relative position of frames. Our method does not use any additional sensor and was evaluated on ex vivo porcine data. It achieves translation and orientation errors of 0.449 \pm 0.189 mm and 1.3 \pm 1.5 degrees respectively for the relative pose estimation. In addition, despite the predominant non-linearity motion in our context, our method achieves a good reconstruction with final and average drift rates of 23.11% and 28.71% respectively. To the best of our knowledge, this is the first work to address volume reconstruction in the context of intravascular ultrasound."

• Mardi 26 septembre 2023 : Présentation de Claire Martin

Titre: "Needle Insertion Training Simulator : Performance and Stability at Different Scales"

Résumé: "Needle-based procedures are often performed, in comparison with open surgeries, for procedures such as biopsies or radiofrequency ablation (RFA) of liver tumors, due to their low invasiveness. The complexity of this task makes training more than essential. Numerical simulators are being developed, facing challenges raised by the need of accurate mechanical and rendering models, as well as the importance of real time computations to allow user interaction. This work focuses on contact problems, and is aimed at being integrated into a large-scale simulation, involving a punctured liver and multiple colliding organs. This work starts from a constraint-based contact model [Duriez et al. (2009)], where a complex compliant mechanical system, coupling different contacts, must be solved. An intersection method is considered to efficiently generate needle-tissue interaction constraints, but it may lead to a high number of contact points. In this work, we use the IsoDOFs method [Zeng et al. (2022)] to quickly build the contact problem in spite of the great size of the system to solve, and we propose a modified Gauss Seidel algorithm to improve the efficiency of the constraint resolution. We also propose a complete pipeline for the simulator, splitting the simulation and rendering loops (both haptic and visual) for optimization purposes. Results show great decreases of the computation time related to the generation and resolution of the contact problem from both the IsoDOFs method and the modified Gauss Seidel algorithm on a simple needle insertion simulation. However, the current precision and computation performances of the large scale simulation still need to be improved so as to make the simulator efficient enough to be considered as a training tool. Ongoing work aims at updating constraint directions at high rates during the resolution step so as to stabilize contacts and improve their outcome."

• Mardi 19 septembre 2023 : Présentation de Vladimir Poliakov

Titre: "Towards Advanced Surgical Training for In-Office Hysteroscopy"

Résumé: "Hysteroscopy is a type of gynecological procedure that enables diagnosis and treatment of intrauterine pathologies using a minimally-invasive approach. Recent advancements, especially in optics, made it possible to perform this procedure in an ambulatory, outpatient setting without anaesthesia. Yet, such approach introduces additional challenges to gynecologists, as the surgical instruments have to pass through the cervix, a narrow curved canal. This process can be traumatic and even cause tissue damage or rupture. Current training curricula mainly focus on general minimally-invasive surgical skills or on inpatient hysteroscopy.This presentation will be threefold. First, I would like to present the solution we designed for surgical simulation development based on the SOFA physics engine and Filament PBR engine. Second, I would like to present the haptic device we designed for translumenal endoscopy. Finally, I will speak of the designed training system and the validation study that we conducted."

• Mardi 12 septemble 2023 : Présentation de François Leconte

Titre: "Enhancing Fluoroscopy-Guided Interventions: a Neural Network to Predict Vessel Deformation without Contrast Agents"

Résumé: "We present a method for estimating, in real-time, a 3D displacement field from a single X-ray image. Our approach uses a fully convolutional network architecture to solve the associated inverse problem. Supervised learning is performed on synthetic data, using Digitally Reconstructed Radiographs as input and displacement fields as output. We use randomized gaussian kernels to produce a synthetic training dataset with displacement fields that are smooth and diffeomorphic. Contrary to other 2D-3D registration methods, our novel data generation approach doesn't rely on statistical motion model, allowing us to predict deformations that are unrelated to breathing motion. Our method is thus well suited for the prediction of the deformation of the vessels when arbitrary deformations are present and no contrast agent is injected. We report a 2D accuracy of 2.7 ± 1.9 mm on synthetic data."

• Mardi 27 juin 2023 : Présentation de Thomas Wahl

Titre: "Closed-loop neurostimulation for the treatment of schizophrenia"

Résumé: "Mental disorders are among the top most demanding challenges in world-wide health. A large number of mental disorders exhibit pathological rhythms, which serve as the disorders characteristic biomarkers. Neurostimulation techniques have been developed to target these pathological rhythms and provide therapeutic interventions. However, current neurostimulation protocols often rely on open-loop approaches, where the stimulation parameters are predefined and independent of the patient's real-time brain state. In this work, we propose a novel fully adaptive closed-loop neurostimulation setup that dynamically adjusts the power spectral density (PSD) of brain activities based on a user-defined PSD. Our approach utilizes a non-parametric brain model estimated from observed data and considers conduction delays in the feedback loop between brain activity measurement and stimulation. We specifically focus on pathological alpha and gamma rhythms associated with psychosis and demonstrate the effectiveness of our method through numerical simulations of neural population and cortico-thalamic loop models. Our findings highlight the potential of closed-loop neurostimulation in improving the treatment of mental disorders by precisely modulating brain activity patterns."

• Mardi 13 juin 2023 : Présentation de Kiwon Um

Titre: "Machine Learning for Visual and Physical Realism"

Résumé: "Solving partial differential equations (PDEs) is a crucial task in all scientific and engineering disciplines. Recently, machine learning techniques have been showing their great capacity for a variety of PDE problems in improving conventional numerical solvers. In this talk, the speaker will discuss two studies of machine learning methods that address the limitation of conventional numerical PDE solvers. First, a neural network model that produces under-resolved splash effects of liquid dynamics will be discussed. Second, the speaker will discuss different learning approaches that find complex correction functions of iterative PDE solvers to reduce numerical errors."

• Mardi 28 mars 2023 : présentation de Elias Cueto

Titre: "Thermodynamics-informed neural networks"

Résumé: "Thermodynamics could be seen as an expression of physics at a high epistemic level. As such, its potential as an inductive bias to help machine learning procedures attain accurate and credible predictions has been recently realized in many fields. We review how thermodynamics provides helpful insights in the learning process. At the same time, we study the influence of aspects such as the scale at which a given phenomenon is to be described, the choice of relevant variables for this description or the different techniques available for the learning process.

We review how thermodynamics can be considered as a high epistemic level inductive bias for techniques of machine learning physical phenomena. Starting with methods imposing conservation laws (particularly, in Hamiltonian or Lagrangian settings) we analyze how the principles of thermodynamics can also be used advantageously for dissipative phenomena. Beginning with methods that simply relax the requirements of conservation and/or equivariance, we see how thermodynamic formalisms can be developed that ensure by construction, either in soft or hard ways, the fulfillment of the principles of thermodynamics: conservation of energy in closed systems and non-negative entropy production. In any case, the main advantage of these approaches is that the resulting accuracy of these methods is considerably improved, or otherwise, the amount of necessary data is drastically minimized."

• Mardi 14 mars 2023 : présentation de Nicola Zotto sur les "Graph Neural Network"

• Mardi 14 février 2023 : présentation de Yu Boyang de 3 articles

1. Progressive Simulation for Cloth Quasistatics: a new forward simulation method for efficient preview of cloth quasistatics on exceedingly coarse triangle meshes with consistent and progressive improvement over a hierarchy of increasingly higher-resolution models. At each coarse-level, the solution is biased with efficiently computed shell forces and energies evaluated on the finest-level model, then each finer-level solve is initialized with safe prolongations of the converged solutions from the prior level’s coarser model.

2. Motion Guided Deep Dynamic 3D Garments: a learning setup that produces dynamic and plausible 3D garment geometry in an iterative roll-out way. A generative space of garment geometries is learned and then mapped to capture the motion-dependent dynamic deformations, conditioned on the previous state of the garment as well as its relative position to the underlying body. The garment dynamics, following the input character motion, are predicted per frame as displacements in the canonical space and then brought to the global space with frame-dependent skinning weights. Any remaining per-frame collisions are corrected by residual local displacements.

3. Neural Cloth Simulation: a general framework for the garment animation problem through unsupervised deep learning inspired by physically-based simulation, the existing optimization scheme for motion from simulation-based methodologies is adapted to deep learning.

• Mardi 31 janvier 2023 : Présentation de Pablo Alvarez

Titre: "Towards a mechanical lung model for video-assisted thoracoscopic surgery."

Résumé: "The resection of pulmonary nodules through video-assisted thoracoscopic surgery is one of the most important surgical procedures for the diagnosis and treatment of early lung cancer. During surgery, pulmonary nodules need to be localized before their resection. However, this is no easy task as the lung undergoes large deformation as a result of a pneumothorax (lung deflation), which is induced at the beginning of surgery. A comprehensive model of lung behavior for VATS may not only provide guidance for nodule localization during surgery, but also help in the planning of the intervention. In this presentation, I will describe our efforts towards building such a model, combining techniques from nonrigid image registration and biomechanical modeling."

• mardi 17 janvier 2023 : Présentation de Guglielmo Scovazzi (USA)

Titre: "The Shifted Boundary / Shifted Fracture Method for Computational Mechanics"

Résumé: "Embedded/immersed/unfitted boundary methods obviate the need for continual re-meshing in many applications involving rapid prototyping and design. Unfortunately, many finite element embedded boundary methods (cutFEM, Finite Cell Method, etc. ) are also difficult to implement due to: (a) the need to perform complex cell cutting operations at boundaries, (b) the necessity of specialized quadrature formulas, and (c) the consequences that these operations may have on the overall conditioning/stability of the ensuing algebraic problems. We present a new, stable, and simple embedded boundary method, named “Shifted Boundary Method” (SBM), which eliminates the need to perform cell cutting. Boundary conditions are imposed on a surrogate discrete boundary, lying on the interior of the true boundary interface. We then construct appropriate field extension operators by way of Taylor expansions, with the purpose of preserving accuracy when imposing the boundary conditions. We demonstrate the SBM in applications involving solid and fracture mechanics; thermomechanics; CFD and porous media flow problems. In the specific case of fracture mechanics, the SBM takes the name of Shifted Fracture Method (SFM), which can be thought of a method with the data structure of classical cohesive fracture FEM but with the accuracy and mesh-independence properties of XFEM. We show how the SFM has superior accuracy in capturing the energy released during the fracture process and how the method can be combined with phase-field approaches to simulate crack branching and merging."

• lundi 14 novembre 2022 : réunion d'équipe à 11h et présentation de prof. Wohn (KAIST)

Titre: "From Cave Paintings to Metaverse"

• lundi 17 octobre 2022 : réunion d'équipe à 11h et présentation de Nora Hagmeyer

Titre: "Towards Modeling Balloon Angioplasty using Geometrically Exact Beam Theory"

• mardi 27 septembre 2022 : réunion d'équipe à 11h

Titre: "Diff-Net: Image Feature Difference Based High-definition Map Change Detection For Autonomous Driving."

Abstract: "Up-to-date High-Definition (HD) maps are essential for self-driving cars. To achieve constantly updated HD maps, we present a deep neural network (DNN), Diff-Net, to detect changes in them. Compared to traditional methods based on object detectors, the essential design in our work is a parallel feature difference calculation structure that infers map changes by comparing features extracted from the camera and rasterized images. To generate these rasterized images, we project map elements onto images in the camera view, yielding meaningful map representations that can be consumed by a DNN accordingly. As we formulate the change detection task as an object detection problem, we leverage the anchor-based structure that predicts bounding boxes with different change status categories. To the best of our knowledge, the proposed method is the first end-to-end network that tackles the high[1]definition map change detection task, yielding a single stage solution. Furthermore, rather than relying on single frame input, we introduce a spatio-temporal fusion module that fuses features from history frames into the current, thus improving the overall performance. Finally, we comprehensively validate our method’s effectiveness using freshly collected datasets. Results demonstrate that our Diff-Net achieves better performance than the baseline methods and is ready to be integrated into a map production pipeline maintaining an up-to-date HD map."

• mardi 12 juillet 2022 : exposé de Valentina Scaponi à 11h

Titre : "Autonomous endovascular catheter navigation"

Résumé : "Endovascular diseases are one of the leading causes of death overall and among all procedures, endovascular interventions are one of the most performed. Only in the USA, this procedure is carried out more than one million times per year, involving different districts of the body, like the heart and the brain, the kidneys and the liver, etc. This procedure involves the insertion of thin tube devices, called guidewires and catheters, inside the patient’s vascular tree. These devices are then navigated toward the target location where they are used to treat or diagnose different types of pathologies. All along the procedure, fluoroscopic images are acquired to let the surgeon see the vessel’s geometry and to allow him to perform the navigation. To acquire these images, both the patient and the surgeon are exposed to x-radiations, which can cause cell damage potentially leading to the development of tumoral masses. The aim of this project is to develop a control algorithm, able to autonomously navigate both the catheter and the guidewire, reducing or even eliminating the exposure to x-radiations thanks to the use of FBGs sensors. This control algorithm will then be used as initial policy of a reinforcement learning algorithm."

• mardi 21 juin 2022 : réunion permanent 10h et exposé de Diwei Wang + réunion d'équipe à 11h

Titre : "Dementia detection in gait videos based on 3D human pose estimation"

Résumé : "Dementia with Lewy bodies and Alzheimer’s disease are two common neurodegenerative diseases among elderly people. They often cause motor impairments such as tremor at rest, rigidity, bradykinesia, and postural instability. Gait analysis is frequently used in clinical applications to detect these anomalies. However, assessments relying on wearable sensors are costly, and sometimes intrusive. Commercial 3D motion analysis systems require carefully calibrated cameras to collect multi-view video data, and are thus not practical. Therefore, we are focused on estimating dementia type and severity using monocular gait videos only. By adopting anomaly mesure-dependent feature learning, and considering the problem of data imbalance, we propose a two-phase training mechanism based on a 3D human pose estimation model. In this presentation, I will explain the approach, present our gait dataset, and show the current problems in modeling temporal correlations between frames."

• mardi 24 mai 2022 : exposé d'Emmanuel Franck (IRMA) à 11h

Titre : "Scientific machine learning: some applications"

Résumé : "Scientific computing, at the interface between numerical analysis, PDE and HPC, has been used for many years to build high-performance simulation codes for physics, biology and medicine. Recently, a new branch, called scientific ML, has emerged. It allows to couple standard scientific computing approaches with deep learning methods. In this presentation, after a quick general presentation, we propose to illustrate this concept through several examples. In a first step, we will focus on the design of reduced models for PDEs from physics. Then we will propose examples concerning the optimization of numerical methods for PDEs. Finally, a control problem from biology will be introduced."

• mardi 17 mai 2022 : réunion d'équipe et exposé de Pierre Galmiche à 11h

Titre : "Towards a modelization of breast evolution during radiotherapy"

Résumé : "Breast cancer is usually treated using radiotherapy after a conservative surgery.

Current methods used to irradiate patients are based on the hypothesis that breast shape and volume don't change during therapy.
This hypothesis has been questioned by some radiotherapists, observing volume and surface changes on patients.
Knowing this, we worked on an approach to track the evolution of shape and volume of breasts during radiotherapy using clinical trial data from the ICANS institute.
During this presentation, I will present our approach, show our results concerning the partial shape matching problem and highlight some difficulties encounted while working on real data."

• mardi 19 avril 2022 : réunion d'équipe et exposé de Boyang Yu à 11h

Titre : "Differentiable Cloth Simulation and Inverse Simulation"

Résumé : "Cloth simulation attracts the attention of many researchers because of its numerous applications in 3D content production such as movies and games. However, classic cloth animation requires expertise and practice, sometimes it may take many attempts to achieve the desired effect. This makes the idea of differentiable cloth simulation appealing. With a proper loss function and by applying gradient-based optimization methods, It can enable automatic control of physical systems or parameter estimation. In this presentation, we will see how to make a physical cloth simulation pipeline differentiable so that the gradients could be backpropagated through it."

• mardi 5 avril 2022 : réunion permanent 10h, réunion d'équipe et exposé de Guillaume Mestdagh à 11h

Titre : "An optimal control problem for elastic registration and force estimation in augmented surgery"

Résumé : "The nonrigid alignment between a pre-operative biomechanical model and an intra-operative observation is a critical step to track the motion of a soft organ in augmented surgery. While many elastic registration procedures introduce artificial forces into the direct physical model to drive the registration, we propose in this paper a method to reconstruct the surface loading that actually generated the observed deformation. The registration problem is formulated as an optimal control problem where the unknown is the surface force distribution that applies on the organ and the resulting deformation is computed using an hyperelastic model. Advantages of this approach include a greater control over the set of admissible force distributions, in particular the opportunity to choose where forces should apply, thus promoting physically-consistent displacement fields. The optimization problem is solved using a standard adjoint method. We present registration registration results with experimental phantom data showing that our procedure is competitive in terms of accuracy. In an example of application, we estimate the forces applied by a surgery tool on the organ. Such an estimation is relevant in the context of robotic surgery systems, where robotic arms usually do not allow force measurements, and providing haptic feedback remains a challenge."

• mardi 22 mars 2022 : réunion d'équipe et exposé de François Lecomte à 11h

Titre : "CNN-based Diffeomorphic Mapping for Real-Time Deformable 2D-3D registration"

Résumé : "We present a method for estimating, in real-time, a 3D displacement field from a single fluoroscopic image. Our approach uses a fully convolutional network architecture to solve the associated inverse problem. Supervised learning is performed on synthetic data, using Digitally Reconstructed Radiographs as input and displacement fields as output. These displacement fields are generated using a diffeomorphic mapping framework to enforce invertibility and smoothness. This also allows our method to be generic and independent of a particular organ deformation or surgical procedure, although it can be tuned to follow the key characteristics of a specific scenario. Results show that our method can estimate the 3D displacement field with an average accuracy of 0.24 ± 0.29mm, at a frame rate of about 20 frames persecond.

• mardi 8 mars 2022 : réunion permanent 10h et exposé de Saurabh Deshpande (Luxembourg) à 11h

Titre : "Probabilistic Deep Learning for Real-Time Large Deformation Simulations"

Résumé : "In this work, we propose a probabilistic deep learning surrogate framework that is capable of accurately predicting non-linear deformations of bodies together with the predictions’ uncertainties. The framework directly takes the Finite Element nodal forces at the neural network input to give nodal displacements at its output. The probabilistic part of the framework is based on a dedicated Variational Inference formulation, thanks to which we are not only able to efficiently capture uncertainties related to noisy data, but we have also knowledge about the model uncertainties—which is especially important in regions not well supported by the data (e.g., the extrapolated region)."

• mardi 15 février 2022 : réunion permanent 10h et exposé à 11h de Leo Nouveau

Titre : "High order embedded strategy for elliptic PDEs in mixed form with the Shifted Boundary Method."

Résumé : "This presentation will be divided in two parts. First, we will see the main features of the Shifted Boundary Method, introduced by A. Main and G. Scovazzi [1]. This method aims at imposing the boundary condition known on an immersed boundary, onto a surrogate one, composed by edges/surfaces of the underlying mesh. To account for the discrepancy between the “true” and “shifted” boundaries, the method modifies the imposed boundary value using Taylor expansions. This allows to preserve the accuracy of the underlying finite element scheme. In a second part dedicated to elliptic PDEs in mixed form, we will investigate how on P1 elements, an overall second order accuracy can be recovered for both Neumann and Dirichlet conditions. This is done using the enrichment of the primary variable [2], and also allows to obtain a third order accurate scheme when only Dirichlet conditions are embedded [3]. This part will be concluded by on-going work, with T. Carlier, H. Beaugendre and M. Colin, for an application to the Stefan problem with moving front."

• mardi 25 janvier 2022 : réunion d'équipe 10h, exposé à 11h de Pierre Mollo

Titre : "Modeling cerebral venous blood flows"

Résumé : "Several factors make venous blood flows much less studied than their arterial counterparts. For example, the venous tree structure presents great variation from one individual to another: preponderance of certain structures, asymmetry, absence of other structures. This makes a generic study very complex. However, we will see in this presentation that by limiting ourselves to cerebral blood flows, we can empirically identify groups of individuals. Moreover, the restriction to the intracranial compartment allows us to use the hypothesis of incompressibility of venous blood vessels. Thus we will see and compare several models for these flows but also their possible extensions. Finally, we will discuss the possibility of using reduced models and the associated applications."

• mardi 9 novembre 2021 : Muhammad Sajjad

Titre : "Efficient Deep Learning Methods for IoT Applications: Current Challenges and Future Directions"

• mardi 26 octobre 2021 : Eligiusz Postek

Titre : "A concept of a coupled agent-stress mechanical model of a tissue"

Résumé : "The physical environment of living cells and tissues, and more particularly their mechanical interaction with it, plays a crucial regulatory role in their biological behavior such as cel differentiation, apoptosis, proliferation, tissue growth, remodeling, wound healing, etc. We will use a concept of coupling of the Agent Based Modelling (ABM) and mechanical modeling. The latter will be applied to single cell models and their colonies. Therefore, it is important to evaluate state of stress in the growing, evolving tissue. To do this, a model of a single cell is necessary as well. The single cell model should consist of the cytoskeleton, cytoplasm, nucleus and cortex."

• mardi 31 août 2021 : Boyang Yu et zhiyu Zheng

Titre de Boyang Yu : "SOTA cloth modeling and its use in motion simulation"

Résumé de Boyang Yu : "The animation of digital humans in clothing has numerous applications in 3D content production. However, cloth modeling requires editing the garment shape in 2D, manually placing, and fine-tuning to achieve final results. This pipeline does need expertise even for an experienced 3D animator. To facilitate cloth modeling, some data-based work has been proposed to learn efficient approximate models, the application of deep learning in recent years has promoted it further. Besides, the reconstruction of humans in clothing is also a topic of interest to researchers. By fitting 3D human & cloth models to a sequence of 3D scans or a video, a more tidy sequence of meshes could be retrieved. In this context, we would like to propose an efficient yet independent of body modeling of cloth, it should be compatible with existing cloth simulator and well parameterized for learning."

Titre de zhiyu Zheng : "On the differentiable cloth simulator for inverse problem"

Résumé de zhiyu Zheng : "This project is to explore a differentiable cloth simulator to see its performance for inverse problem like motion control and external force estimation. In our project, we have studied the differentiable cloth simulator coupled with deep neural networks. Firstly, we have tested the performance of the simulator with motion control task, and we found that the simulator coupled with the neural networks can be trained to execute different tasks of throwing the piece of cloth into the basket at different locations. Our result only provides a view of the possibility for more extended applications, and it could be further optimized with more trainings. Then we test its performance for dropping force estimation, we found that given a sequence of cloth mesh frames, the simulator can well determine the dropping force according to their deformation and movement. This is important and could be potentially applied for more complicated external force estimation like the force applied by the human body on the cloth."

• mardi 13 juillet 2021 : Anne-Sarah Debly

Titre : "Autonomous endovascular navigation using reinforcement learning"

Résumé : "Catheters are mainly used to treat cardiac diseases like strokes and heart attacks but navigating through the vascular tree is a hard task for the surgeon. Today, he can follow the position of the catheter in the patient's vessels thanks to fluoroscopic imaging, but it gives only 2D images, with a complex visualization of the vessels. For these reasons, we propose to develop an estimation of the 3D shape of the catheter in 3D thanks to FBG sensors placed on an optic fiber to retrieve the 3D shape of the device. Then, we would like to develop a closed-loop system based on depp reinforcement learning to make possible the autonomous navigation of the catheter. By combining knowledge of the 3D shape of the catheter from the sensors, and a deep learning algorithm to choose the best path to the target, a robot could control the catheter tip."

Enregistrement : bbb

Slides : pdf

• mardi 29 juin 2021 : Virginie Marec

Titre : "Integration of a force sensor for online estimation of the behaviour of the system and parametrization of the real-time simulation during robotic needle insertion."

Résumé : "This internship focuses on real-time updates of a biomechanical model of a needle insertion procedure based on force sensor data. In fact, a force sensor can be integrated at the base of a needle in order to measure in real time the force applied to the latter during its insertion. These data will then make it possible to achieve 2 objectives: -Detect layer changes when inserting the needle in a heterogeneous gel (skin, fat, muscle, liver). -Estimate the frictional forces applied along the needle. In parallel, a finite element model is implemented on SOFA. We want to obtain a realistic simulation that correctly models the behavior and the deformations of the tissues during the insertion of the needle. The final objective of this project is to create a method which, from the values ​​obtained from a force sensor, is able to deduce the state of the system (i.e. in which tissue is the needle and how friction acts on it). The parameters of the SOFA model can then be adjusted in real time so that the simulation best matches reality."

Enregistrement : bbb

Slides : pdf

• mardi 22 juin 2021 : Josephine Riedinger

Titre : "Closed-loop Transcranial Electric Stimulation of Neural Networks in a Rodent Model of Psychotic Transition"

Résumé : "Schizophrenia is a psychotic disorder characterized by a loss of contact of the patient with reality. The prodromal phase is the period in which some symptoms happened and announced the onset of the disease. Today, there is an increasing interest in preventing the psychotic transition of patients-at-risk (in the prodromal phase) to the chronic psychotic phase of the disease. Interestingly, it was shown that patients in this prodromal phase shown abnormal cerebral oscillations that can be used as biomarkers. Hence, we can ask if the normalization of these "oscillopathies" could delay, or even prevent, the psychotic transition and so, the entry of the patients in the chronic phase of the disease. One way could be the application of a neuromodulation technique, such as the promising Transcranial Electrical Stimulation (TES). In this pre-clinical project, we would like to explore if a closed-loop TES application can normalize the oscillopathies found in patients. For this purpose, electrophysiological experiments are conducted in an animal model of ketamine-induced psychotic transition. Complementary, a model of the brain network of interest is designed and an adapted control system is researched. An extended Kalman filter will allow the model observer to take into account real brain activity for the computation of the predictions of states and observations. Advancements and first results are encouraging and will be presented."

Enregistrement : bbb

• mardi 15 juin 2021 : Thomas Wahl et Philippe Pincon

Titre de Thomas Wahl : "Effects of drugs and neurostimulation on gamma-oscillations in neural networks."

Résumé de Thomas Wahl : "Psychosis in schizophrenia is known to be correlated with a stronger activity in the frequency range 25-60 Hz, corresponding to γ-oscillations in electroencephalograms (EEGs). Neurostimulation can be used to attenuate this effect. By using a mean-field description derived from a network of interacting excitatory and inhibitory neural population with additive noise input, we can apply linear response theory to study analytically the properties of γ-oscillations, to explain the emergence of noise driven oscillations, also known as quasi-cycles. We can use our understanding of the properties of quasi-cycles to show in details the influence of our model parameters on the properties of the γ-oscillations."

Titre de Philippe Pincon : "Effects of drugs and neurostimulation on gamma-oscillations in neural networks."

Enregistrement : bbb

• mardi 8 juin 2021 : Rongrong Liu

Titre : "Wearable Sensor Technology for Individual Grip Force Profiling."

Résumé : "Wearable biosensor systems with transmitting capabilities represent innovative technology developed to monitor exercise and other task activities. This technology enables real-time, convenient, and continuous monitoring of a user’s behavioral signals, relative to body motion, body temperature and a variety of biological or biochemical markers, like individual grip force, which is studied here. To achieve this goal, a four-step pick-and-drop image-guided robot-assisted precision task has been designed using a wearable wireless sensor glove system. The spatio-temporal grip force profiling is analyzed on the basis of thousands of individual sensor data collected from the twelve locations on the dominant and non-dominant hands of each of the three users in ten successive task sessions. Statistical comparison has shown specific differences between the grip force profiles of individual users as a function of task skill level and expertise."

Enregistrement : bbb

Slides : pdf

• mardi 25 mai 2021 : Pierre Galmiche

Titre : "The Functional representation of the shapes."

Résumé : "Breast cancer can be treated using radiotherapy after a conservative surgery. The current method used to irradiate patients is based on the hypothesis that the breast shape and volume don't change across radiotherapy sessions. This hypothesis has been questioned by some radiotherapists, observing volume and surface changes during therapy. Knowing this, we want to track the deformation across radiotherapy using clinical trial data from the ICANS institute. To answer this problematic we chose to focus on the Spectral Representation of the shapes to solve the Shape Matching Problem."

Enregistrement : bbb

Slides : pdf

• mardi 18 mai 2021 : Paul Baksic

Titre : "Shared control strategy for �needle insertion into deformable tissue using inverse Finite Element simulation"

Résumé : "We have previously proposed a fully automated strategy for percutaneous procedures. This algorithm is relying on a FEM simulation used to derive the Jacobian linking the needle base motion to its tip motion relative to the target (the tumor) while taking into account needle-tissue interaction. This jacobian is used to compute the motion of the robot holding the needle needed to reach the target. To compute it, a FE model of the liver is registered using 3D positions of a small set of points in the liver measured in real-time. But to do so, it is assumed that those 3D positions can be measured in real-time, which is currently not possible. In our recent work, we propose a more realistic shared-control framework where only 2D information of those points, coming from a C-ARM, are used to register the model. This introduces higher registration errors that cannot be compensated by the automatic algorithm. That is why we have proposed a new shared control strategy for needle insertion into soft tissue. It consists in leaving the decision-making part (when to insert the needle, and where should the tip go inside of the tissue) to the practitioner. This is done through a haptic interface, throughout which the user controls the target of the automatic algorithm inside of the tissue. In addition, the user is guided on a predefined path using virtual fixtures. Alongside this haptic interface, augmented 2D fluoroscopic images are provided to the practitioner. This allows leaving the complex motion of the needle base needed to take the needle-tissue interaction into account to the automated algorithm, while the user can focus more on the important part which is, where to place the needle tip. This method an the results of a user study published at ICRA2021 are presented here. In addition, efforts are made to go from a simulated environment to a real phatom trial. This is still a work in progress and the last advances are presented here. "

Enregistrement : bbb

• mardis 4 et 11 mai 2021 : Cedric Bobenrieth

Titre : "Modélisation Géométrique par Croquis"

Résumé : "De nos jours, la modélisation 3D est omniprésente, cependant les outils modernes pour créer des modèles 3D sont complexes et requièrent beaucoup de temps. A contrario, l’esquisse est un moyen naturel de communiquer rapidement des idées, ainsi une méthode permettant la reconstruction automatique d’objets 3D à partir d’un croquis simplifierait ce processus. Cette méthode devrait résoudre deux problèmes : le calcul des parties cachées de la forme dessinée et la détermination des coordonnées 3D à partir des données 2D du croquis. Dans cette présentation, je vous parlerais de deux nouvelles approches qui visent à surmonter ces problèmes. La première se sert d’a priori et d’une base de données préexistantes pour permettre la reconstruction 3D automatique de fleurs à partir d’un seul croquis selon n’importe quel angle de vue. La seconde permet la reconstruction de tout type d’objets, sans limitations, en utilisant un style de dessin plus informatif et en étant guidée par l’utilisateur."

Enregistrement : bbb

Slides : pdf

• mardi 27 avril 2021 : Dawood Al Chanti

Titre : "IFSS-Net: Interactive Few-Shot Siamese Network for Faster Muscles Segmentation and Propagation in 3-D Freehand Ultrasound"

Résumé : "We present an accurate, fast and efficient method for segmentation and muscle mask propagation in 3D freehand ultrasound data, towards accurate volume quantification. To this end, we propose a deep Siamese 3D Encoder-Decoder network that captures the evolution of the muscle appearance and shape for contiguous slices and uses it to propagate a reference mask annotated by a clinical expert. To handle longer changes of the muscle shape over the entire volume and to provide an accurate propagation, we devised a Bidirectional Long Short Term Memory module. To train our model with a minimal amount of training samples, we propose a strategy to combine learning from few annotated 2D ultrasound slices with sequential pseudo-labeling of the unannotated slices. To promote few-shot learning, we propose a decremental update of the objective function to guide the model convergence in the absence of large amounts of annotated data. Finally, to handle the class-imbalance between foreground and background muscle pixels, we propose a parametric Tversky loss function that learns to adaptively penalize false positives and false negatives. We validate our approach for the segmentation, label propagation, and volume computation of the three low-limb muscles on a dataset of 44 subjects. We achieve a dice score coefficient of over 95% and a small fraction of error with 1.6035%."

Enregistrement : bbb

• mardi 20 avril 2021 : Mathieu Naudin

Titre : "Virtual biopsy"

Résumé : "Virtual biopsy is a real societal issue from screening to diagnosis through patient follow-up. It is defined as allowing, from a set of data, to propose a diagnosis and to link it to an uncertainty. The goal is to one day consider the disappearance of physical biopsy for obvious reasons: invasive, traumatic and sometimes dangerous or unfeasible. This method requires an important mix of expert knowledge and artificial intelligence. It uses formatted measurements as an entry point, and neural networks specific to the desired diagnosis and thus respond to the medical problem as a whole. We will see the implementation of this approach through some of my thesis work as well as my recent research on the premises of the following axes: digital twin, lesion microenvironment and deployment in real conditions."

• mardi 13 avril 2021 : Guillaume Mestdagh

Titre : "An optimal control approach for surface-matching in augmented surgery"

Résumé : "Augmented surgery consists in providing in real-time a 3D view of an organ during a surgical intervention. In this context, the displacement field in the organ is reconstructed from partial data. We propose an optimal control approach for this problem, involving an elastic model describing the organ's response to surface loadings and a functional measuring the discrepancy between the current displacement field and available data. In this formulation, we try to reconstruct a physically plausible surface loading field rather than create artificial forces to generate a displacement. In this presentation, we introduce the optimal control formulation and discuss its advanges, and then we show a numerical example on a toy problem."

Slides : pdf

Enregistrement : bbb

• mardi 30 mars 2021 : Ziqiu Zeng

Titre : "Method of isolating dofs on dealing with contact equtations"

Résumé : "The construction of compliance matrix is usually a main obstacle in real-time FE simulation with interactions between objects. Recently we have developed a new GPU-based method to solve the contact equations using LDL factorization and nested dissection. The approach significantly reduces the computation cost for contact resolution. In the meeting I will present our cmputation strategy as well as our latest results."

• mardi 23 mars 2021 : Hyewon Seo

Titre : "Dynamic skin deformation prediction by recurrent neural network"

Résumé : "Skin dynamics contributes to the enriched realism of human body models in rendered scenes. Traditional methods rely on physics-based simulations to accurately reproduce the dynamic behavior of soft tissues. Due to the model complexity and thus the heavy computation, however, they do not directly offer practical solutions to domains where real-time performance is desirable. The quality shapes obtained by physics-based simulations are not fully exploited by example-based or more recent data-driven methods neither, with most of them having focused on the modeling of static skin shapes by leveraging quality data. To address these limitations, we present a learning-based method for dynamic skin deformation. At the core of our work is a recurrent neural network that learns to predict the nonlinear, dynamics-dependent shape change over time from pre-existing mesh deformation sequence data. Our network also learns to predict the variation of skin dynamics across different individuals with varying body shapes. After training the network delivers realistic, high-quality skin dynamics that are specific to a person in a real-time course. We obtain results that significantly saves the computational time, while maintaining comparable prediction quality compared to state-of-the-art results."

Slides : pdf

Enregistrement : bbb

• mardi 16 mars 2021 : François Lecomte

Titre : "Recalage 3D/2D grâce au Deep Learning sur des données synthétiques"

Résumé : "Cette présentation parlera du recalage de données CT sur des images fluoroscopiques. Je présenterai notre processus de génération de données synthétiques (DRRs) à partir d’un CT pré-opératoire. J’exposerai ensuite brièvement l’architecture du réseau et le processus d’apprentissage. Après un résumé du processus complet, je présenterai nos résultats sur le dataset de validation synthétique. Je conclurai sur les limites actuelles et les prochaines étapes dans le développement de la méthode."

Slides : pdf

Enregistrement : bbb

• mardi 9 mars 2021 : Robin Enjalbert

Titre : "Automatic catheter navigation through deep reinforcement learning"

Résumé : "For this meeting, I will introduce you the basics of Deep Reinforcement Learning as well as a specific algorithm, Deep Q-Network. Then, I will present my work on Deep Reinforcement Learning for the navigation of endovascular catheters."

Slides : pdf

Enregistrement : bbb

• mardi 2 mars 2021 : Axel Hutt

Titre : "Models of drug and stimulation impact on neural populations"

Résumé : "The talk shows recent mathematical and numerical results on the drug impact and stimulation impact on brain models. The effects are considered in the context of cognition, general anaesthesia and mental disorders."

Slides : pdf

Enregistrement : bbb

• mardi 23 février 2021 : Sergei Nikolaev

Titre : "Parameters estimation using Kalman filters for predictive simulation"

Résumé : "In this talk, we are going to talk about Kalman filters. Firstly, there will be a brief introduction. Then we will see several options to reduce the order of the filter, in order to obtain the data assimilation process close to real-time. Finally, several examples will be presented."

Slides : pdf

Enregistrement : bbb

• mardi 16 février 2021 : Jean-Nicolas Brunet

Titre : "Creating python bindings with pybind11 and SofaPython33"

Résumé : "During this meeting, we will go through the basics of creating packages and modules in python, and how they are automatically found by python. We will then learn how to create packages and modules in C++ with pybind11. Finally, we will go through the bindings of a simple SOFA plugin using both pybind11 and SofaPython3. We will see how we can call C++ functions from our components directly in python, how to inherits a c++ class from a python class and how to bind specific SOFA data types in python."

Enregistrement : youtube

Code source du tutoriel : https://github.com/jnbrunet/tutorial_sp3

Exemple SofaOffscreenCamera : https://github.com/jnbrunet/SofaOffscreenCamera

• mardi 2 février 2021 : Michel Duprez

Titre : "φ-FEM: une méthode éléments finis aux frontières immergées sur des domaines définis par une fonction level-set"

Résumé : Les méthodes éléments finis classiques utilisent des maillages qui coïncident avec le bord et les interfaces du domaine sur lequel nous effectuons la simulation numérique. Suivant le type d'éléments du maillage utilisés ou lorsque la géométrie du domaine est trop complexe, une méthode alternative consiste à effectuer les calculs sur un maillage qui ne coïncident pas avec le bord et les interfaces du domaine. φ-FEM appartient à cette classe de techniques et a la particularité de tenir compte des forces externes à l'aide d'une fonction "level-set" qui s'annule au bord. Dans cet exposé, je rappellerai tout d'abord la méthode des éléments finis "classique" et quelles contraintes géométriques doivent satisfaire les maillages. Je présenterai ensuite les différentes méthodes aux frontières immergées précédentes, leurs avantages et leur inconvénients. Enfin, j'introduirai φ-FEM, quel verrous scientifique cette méthode permet de lever et dans quel cadre elle a été développée pour le moment. Je terminerai par quelques simulations numériques.

Slides : pdf