HAL Portal UPHF (Université Polytechnique Hauts-de-France)
Not a member yet
28473 research outputs found
Sort by
Gated Attention-Augmented Double U-Net for White Blood Cell Segmentation
No full textInternational audienceSegmentation of white blood cells is critical for a wide range of applications. It aims to identify and isolate individual white blood cells from medical images, enabling accurate diagnosis and monitoring of diseases. In the last decade, many researchers have focused on this task using U-Net, one of the most used deep learning architectures. To further enhance segmentation accuracy and robustness, recent advances have explored the combination of U-Net with other techniques, such as attention mechanisms and aggregation techniques. However, a common challenge in white blood cell image segmentation is the similarity between the cells' cytoplasm and other surrounding blood components, which often leads to inaccurate or incomplete segmentation due to difficulties in distinguishing low-contrast or subtle boundaries, leaving a significant gap for improvement. In this paper, we propose GAAD-U-Net, a novel architecture that integrates attention-augmented convolutions to better capture ambiguous boundaries and complex structures such as overlapping cells and low-contrast regions, followed by a gating mechanism to further suppress irrelevant feature information. These two key components are integrated in the Double U-Net base architecture. Our model achieves state-of-the-art performance on white blood cell benchmark datasets, with a 3.4% Dice score coefficient (DSC) improvement specifically on the SegPC-2021 dataset. The proposed model achieves superior performance as measured by mean the intersection over union (IoU) and DSC, with notably strong segmentation performance even for difficult images
LoLA-SpecViT: Local attention SwiGLU vision transformer with LoRA for hyperspectral imaging
No full textOur code is available in the following : https://github.com/FadiZidiDz/LoLA-SpecViT-Model.gitInternational audienceHyperspectral image classification remains a challenging task due to the high dimensionality of spectral data, significant inter-band redundancy, and the limited availability of annotated samples. While recent transformer-based models have improved the global modeling of spectral-spatial dependencies, their scalability and adaptability under label-scarce conditions remain limited. In this work, we propose LoLA-SpecViT(Low-Rank Adaptation Local Attention Spectral Vision Transformer), a lightweight spectral vision transformer that addresses these limitations through a parameter-efficient architecture tailored to the unique characteristics of hyperspectral imagery. Our model combines a 3D convolutional spectral front-end with local window-based self-attention, enhancing both spectral feature extraction and spatial consistency while reducing computational complexity. To further improve adaptability, we integrate enhanced Low-Rank Adaptation (LoRA) into attention and projection layers, enabling fine-tuning with over 70% fewer trainable parameters. A novel cyclical learning rate scheduler modulates LoRA adaptation strength during training, improving convergence and generalization. Extensive experiments on four benchmark datasets, WHU-Hi LongKou, WHU-Hi HongHu, Salinas, and QUH-Qingyun, demonstrate that LoLA-SpecViT consistently outperforms state-of-the-art baselines, achieving up to 99.91% accuracy with substantially fewer parameters and enhanced robustness under low-label regimes. The proposed framework provides a scalable and generalizable solution for real-world HSI applications in agriculture, environmental monitoring, and remote sensing analytics. Our code is available in the following GitHub Repository.</div
Architectures bioinspirées ultra-faible consommation pour la localisation et la reconnaissance de sources sonores
No full textBiomimetism is an increasingly widespread approach in various scientific fields. It regularly gives rise to new paradigms and, in recent years, has driven neuromorphic computing and technologies which promise significant advances in the field of information theory and unprecedented energy efficiency. With this approach, artificial spiking systems inspired by the neuronal impulse processing in the brain can process signals from various modalities. In the context of acoustic monitoring of biodiversity, this thesis investigates the potential of an analog neuromorphic technology integrating metal-oxide-semiconductor field-effect transistors operating in the subthreshold regime with ultra-low power (ULP) consumption. Keeping ULP constraints under consideration, bioinspired energy-efficient precomputing tools are designed for sound source localization and recognition and their performances assessed. Firstly, an original interaural time difference (ITD) extractor is modelled after the Hassenstein-Reichardt detector of motion detection, chosen for its low number of neurons, and applied to acoustic signals for estimation of sound sources’ direction of arrival. The ITD extractor is evaluated in simulation on the basis of 2-D and 3-D indoor recordings at distances between 24 cm and 10 m of click-like sounds in particular. A simplified hyperbolic multilateration technique enables the analysis of the ITD extractor’s localization performances, resulting in encouraging azimuth accuracies, in view of its potential ULP consumption, of 73.9% (±2.5°) and 77% (±5°) between 1 m and 3 m for click-like sounds. Then, with the aim to address multisource scenarios, a detector of temporal characteristics inspired by the calling song recognition mechanism of female field crickets is designed and successfully implemented on chip using the subthreshold neuromorphic technology. Tested under a probe station with artificial and real-world cricket calling songs, the detector reaches a sub-nanowatt total power consumption in quiet or noisy scenarios with high precision and encouraging recall. Finally, combining multiple instances of these two precomputing tools enables one to envision acoustic source tracking and counting applications.Le biomimétisme est une approche de plus en plus répandue dans les différents domaines scientifiques. Il est régulièrement la source de nouveaux paradigmes et, depuis quelques années, a impulsé le traitement et les technologies neuromorphiques qui portent la promesse d’avancées significatives dans le domaine de la théorie de l'information et une efficacité énergétique sans précédent. Avec cette approche, les systèmes artificiels à impulsions inspirés du traitement neuronal dans le cerveau peuvent traiter des signaux issus de diverses modalités. Dans le contexte de la surveillance acoustique de la biodiversité, cette thèse explore le potentiel d'une technologie neuromorphique analogique intégrant des transistors à effet de champ métal-oxyde-semi-conducteur fonctionnant en régime sous-le-seuil avec une puissance consommée ultra-faible (ULP). En tenant compte des contraintes ULP, des outils de pré-traitement bioinspirés et économes en énergie sont conçus pour la localisation et la reconnaissance des sources sonores et leur performances sont évaluées. Tout d'abord, un extracteur original de différence interaurale de temps (ITD) est modélisé d'après le détecteur de mouvement Hassenstein-Reichardt, choisi pour son faible nombre de neurones, et appliqué aux signaux acoustiques pour estimer la direction d'arrivée des sources sonores. L'extracteur d’ITD est évalué en simulation sur la base d'enregistrements en intérieur en 2D et 3D à des distances comprises entre 24 cm et 10 m, en particulier pour des sons de type clic. Une technique simplifiée de multilatération hyperbolique permet d'analyser les performances de localisation de l'extracteur d’ITD, et qui résulte en des précisions azimutales encourageantes, compte tenu de sa consommation potentielle d'ULP, de 73,9 % (±2,5°) et 77 % (±5°) entre 1 m et 3 m pour les sons de type clic. Ensuite, dans le but de traiter des scénarios multisources, un détecteur de caractéristiques temporelles inspiré du mécanisme de reconnaissance du chant d'appel des criquets femelles a été conçu et intégré sur puce à partir de la technologie neuromorphique sous-le-seuil. Testé sur un banc sous pointes avec des signaux artificiels et des chants d'appel de criquets réels, le détecteur atteint une consommation totale inférieure au nanowatt dans des scénarios calmes ou bruyants, avec une grande précision et un recall encourageant. Finalement, la combinaison de plusieurs instances de ces deux outils de pré-traitement permet d'envisager des applications de suivi et de comptage des sources acoustiques
Identification of the differential and synergic lipotoxic patterns of oleic acid, palmitic acid, and their mixture in 3D HepG2/C3A tissue using liver‐on‐chip technology
No full textInternational audienceThe metabolic dysfunction‐associated steatotic liver disease (MASLD, previously formerly known as non‐alcoholic fatty liver disease, NAFLD) is rapidly expanding worldwide in parrallel with the obesity pandemic. Dietary fatty acids including oleic (OA) and palmitic acids (PA) contribute to the hepatic intracellular triglyceride accumulation, and are therefore thought to play key roles in disease development and progression. Taking advantage of the cutting‐edge organ‐on‐chip technology that mimics the 3D organ dynamic environment, we aimed at investigating the role of OA, PA and a 2:1 OA/PA mixture on the growth and function of the HepG2/C3A, a liver cell line model, over 2 and 7 days. OA supported sustained cell growth, leading to dense 3D tissues, whereas PA and OA exposure did not affect cell proliferation. PA treatment downregulated the GLUT2, INSRA, SREBP1, FASN, mRNA levels indicating a lipid metabolism perturbation in our model. The cell dysfunction caused by OA, PA, and OA/PA was associated with an increase in reactive oxygen species (ROS) production over time. Intracellular lipid monitored by oil red O was higher in cells exposed to OA than in the control ones and cells cultured with PA. Our data confirm the role of fatty acids on the growth and dysfunction of HepG2/C3A cells, and highlight distinct mechanisms through which OA and PA exert their effects
Remarkable electronic and optical properties of HgTe quantum dots
No full textInternational audienceHgTe is a very special material. Semi-metallic in bulk, it can become semiconducting under the effect of quantum confinement. This is well known in the case of quantum wells based on HgCdTe/HgTe/HgCdTe heterostructures with infrared optoelectronic properties at the heart of current applications. However, it has been known for some years that the electronic structure of these quantum wells can exhibit non-trivial topological properties. In this talk, we look at the case of colloidal nanocrystals (quantum dots), the synthesis of which is now well understood. We show that topological effects strongly influence the optoelectronic properties of these HgTe quantum dots. As a result, these properties differ considerably from those of quantum dots of conventional semiconductors such as CdTe. We show that these original effects open up interesting prospects for optoelectronic applications from infrared to THz
Convex Conditions for Observer Design in Nonlinear Continuous-Time Systems Using a Spatial Discretization Procedure
No full textInternational audienceThis paper proposes convex conditions for the observer design of nonlinear continuous-time systems. A broad class of nonlinear systems can be tackled by the proposed technique. A spatial discretization is employed, and an approximate model is obtained within the error matrices that measure the difference between the nonlinear system and the approximated one. The conditions are formulated as parameter-dependent matrix inequalities and ensure that the observer can asymptotically follow the states of the original nonlinear system while guaranteeing a bound to the L2-gain from the disturbance input to the estimation error. Numerical experiments are used to illustrate the features of the proposed method
Droplet velocity in both limits of low and high soluble surfactants in a Hele-Shaw cell: experimental and analytical results
No full textInternational audienceThe transport of droplets in microfluidic channels is strongly dominated by interfacial properties, which makes it a relevant tool for understanding the mechanisms associated with the presence of more or less soluble surfactants. In this paper, we show that the mobility of an oil droplet pushed by an aqueous carrier phase in a Hele-Shaw cell qualitatively depends on the nature of the surfactants: the drop velocity is an increasing function of the drop radius for highly soluble surfactants, whereas it is a decreasing function for poorly soluble surfactants. These two different behaviours are experimentally observed by using two families of surfactant with a carbon chain of variable length. We first focus on the second regime, observed here for the first time, and we develop a model which takes into account the flux of surfactants on the whole droplet interface, assuming an incompressible surfactant monolayer. This model leads to a quantitative agreement with the experimental data, without any adjustable parameter. We then propose a model for a stress-free interface, i.e. for highly soluble surfactants. In these two limits, the models become independent on the physico-chemical properties of the surfactants, and should be valid for any surfactant complying with the incompressible or stress-free limit. As such, we provide a theoretical framework with two limits for all the experimental physico-chemical configurations, which constitute the bounds for the droplet mobility for intermediate surfactant solubility
Termal fluctuations effect on crack propagation and decohesion phenomena
No full textInternational audienceFracture propagation and decohesion phenomena are widely present in various physical, biological, and technologicalsystems. The study of these phenomena has applications that extend from classical mechanical sciences (e.g., crack propagationin solids) to emerging fields such as advanced materials, nanotechnology [1, 2], and soft materials [3], includingrubber-like and biomaterials (e.g., cell adhesion and de-adhesion, DNA denaturation) [4, 5]. In many of these fields, thermaleffects may play a crucial role. For instance, the required force for DNA denaturation or for the peeling of a thin filmfrom a substrate typically diminishes as temperature rises. This reduction is often attributed to thermal fluctuations, whichfacilitate the exploration of phase space [6]. In some cases, such effects can lead, as temperature grows, to a spontaneousevolution of fracture and decohesion fronts. Despite extensive experimental studies and numerical simulations exploringthe relationship between thermal fluctuations and mechanical properties, a comprehensive theoretical approach that fullyintegrates temperature effects remains elusive.Here, based on previous studies by the authors [12, 13], we address this gap, with the main goal of rigorously elucidatingthe influence of temperature on decohesion and material failure. One of the challenges in this analysis is to bridge thenano-scale with the meso-scale through a multiscale approach. In the presented work, this paradigm is employed by concurrentlyintroducing a discrete and a continuum model of the mode I fracture. As demonstrated in [12, 13], this approachensures that both the local and global minimizers of the discrete model converge to the equilibria of the mesoscopic modelin the continuum limit, thereby preserving all the essential physical information of the discrete (low-scale) lattice. Thediscrete model lets us incorporate temperature effects through the tools of classical equilibrium statistical mechanics [8, 9],measuring fluctuations at molecular- or nano-scales, whereas the continuous model enables the description of elasticityand energy release rate at the mesoscale.Figure 1 shows the general idea we used to study crack propagation, adhesion, and decohesion using a multiscaleapproach. Starting from the macroscopic description of an elastic body containing a crack subjected to tensile stress,we first consider a discrete model composed of repetitions of identical units at a small scale (Figure 1 a). Applyingdownscaling (Figure 1 b) and c), we first consider a discrete prototypical model composed of perfectly elastic elements(representing the two elastic half planes) connected by elasto-fragile elements (Figure 1 d). We then study the continuumlimit (i.e. the limiting value of the total energy when the number of units of the discrete system tends to infinity) to gaininsight into mesoscopic phenomena. The deduction of continuous, compact formulations is highly desirable both forthe deduction of multiscale approaches and for clearer physical interpretability of the analytical results. Through thisapproach, building on the Griffith energy criterion [7], we are then able to extend the established criterion for fracture anddecohesion to incorporate the influence of thermal fluctuations. The main innovation of our approach involves substitutingtotal mechanical energy with thermodynamic quantities, specifically free energies.The considered prototypical model, extensible in many possible materials and geometric directions at the expensesof physical clarity and analyticality, allows us both to effectively capture all main qualitative features of the fracturephenomenology and to integrate temperature effects. Moreover, all results are deduced in closed analytical form, allowingfor deeper description of the fundamental underlying physical phenomena. In addition, these simple models allow us todescribe either the case where fracture or decohesion propagates from one end, or the configuration where the broken partis inside the system, including scenarios with multiple bubbles in the system (e.g., multiple bubbles in DNA). In the lattercase, we can determine solutions where these bubbles coalesce before failure.Interestingly, our approach unveils a classical critical behavior, with the critical load decreasing according to thelaw sqrt(1 − T/Tc) as the temperature increases [10, 11]. As a consequence of the temperature effects, we prove that incorrespondence to the critical temperature Tc , the system undergoes a phase transition, corresponding to the completerupture without the application of any mechanical load.In analogy with the quasicontinuum framework [14], the proposed description of thermal effects can be incorporatedinto real physical systems by interpreting the model as a representation of the process zone in decohesion and crackpropagation phenomena. This is the objective of our future investigations
Operationalizing selective transparency using progressive disclosure in artificial intelligence clinical diagnosis systems
No full textRevised preprint version submitted to the International Journal of Human-Computer Studies (Elsevier).Final published version available on ScienceDirect at DOI: 10.1016/j.ijhcs.2025.103591Explainable AI (XAI) is critical for clinical decision support systems (AI-CDSS) in healthcare, but current approaches often neglect the usability of explanations from a human-computer interaction (HCI) perspective. We investigate progressive disclosure as a strategy for selective transparency to provide effective explanations without overwhelming users. This paper presents a user-centered design of AI-CDSS interface prototypes that incorporate interactive explanation features (e.g., keyword highlighting of medical terms and interactive causal diagrams) and empathy-oriented nudges (e.g., supportive prompts and icons). We evaluated these prototypes through interviews with medical professionals and students, followed by a user study with general users, to assess their impact on understanding, trust, and satisfaction. Our findings suggest that progressive, on-demand disclosure of explanation details may help users manage information load and better follow the AI’s reasoning process. While several interface features were well received, some elements such as affective cues like emojis elicited skepticism, particularly in clinical contexts, which underscores the importance of context-sensitive design choices