Portail HAL des publications du LIRMM
Not a member yet
13279 research outputs found
Sort by
Développement d’un pipeline standardisé d’apprentissage profond pour la segmentation précise des tumeurs primaires et des métastases péritonéales dans le cancer de l’ovaire sur des images CT.
No full textInternational audienceOvarian cancer, the second most lethal gynecological tumor, is predominantly represented by high-grade serous ovarian carcinoma (HGSOC), which accounts for 70 to 80% of associated deaths. In addition to a poor prognosis (5-year survival < 50%), assessing the extent of the disease remains a major challenge. Deep Learning (DL) approaches applied to the segmentation of ovarian and peritoneal lesions in medical imaging (CT, MRI) have shown promising results. However, questions remain regarding model generalizability, performance differences between 2D and 3D models, and their clinical relevance.We developed a standardized DL pipeline to accurately segment primary and metastatic ovarian lesions. Based on PyTorch Lightning and MONAI, this pipeline includes modules compatible with standard data formats (JSON, NIFTI, Parquet) and offers advanced, extensible models (e.g., ResUnet, ResUNet++) with the option to add attention modules. Standardized loss functions and evaluation metrics, aligned with recent literature recommendations, complete the architecture.A public dataset of 337 annotated CT scans, including 90 validated by clinicians, was used to train and evaluate a binary 2D UNet model on the Jean Zay supercomputer. Initial results demonstrate promising potential for accurate segmentation of ovarian and metastatic lesions (Figure 1, Table 1). This work aims to standardize and improve the robustness of DL models, facilitating their clinical adoption and addressing the specific needs of oncology segmentation.Le cancer de l'ovaire, deuxième tumeur gynécologique la plus mortelle, est majoritairement représenté par le carcinome séreux de haut grade (HGSOC), responsable de 70 à 80 % des décès associés. En plus d’un mauvais pronostic (survie à 5 ans < 50 %), l’évaluation de l’extension de la maladie reste un défi majeur. Les approches de Deep Learning (DL) appliquées à la segmentation des lésions ovariennes et péritonéales sur imagerie médicale (CT, IRM) ont montré des résultats prometteurs. Toutefois, des questions persistent concernant la généralisation des modèles, les différences de performances entre des modèles 2D et 3D, et leur pertinence clinique.Nous avons développé un pipeline DL standardisée pour segmenter les lésions ovariennes primaires et métastatiques avec précision. Basée sur PyTorch Lightning et MONAI, cette pipeline intègre des modules compatibles avec des formats de données standards (JSON, NIFTI, Parquet) et propose des modèles avancés (ex : ResUnet, ResUNet++) extensibles, avec la possibilité d’ajouter des modules d’attention. Des fonctions de perte et des métriques d’évaluation standardisées, suivant les dernières recommandations de la littérature, complètent cette architecture.Un dataset public de 337 CT annotées, dont 90 validées par des cliniciens, a été utilisé pour entraîner et évaluer un modèle UNet 2D binaire sur le supercalculateur Jean Zay. Les résultats initiaux révèlent un potentiel prometteur pour la segmentation précise des lésions ovariennes et métastatiques (Figure 1, Table 1). Ce travail vise à standardiser et renforcer la robustesse des modèles DL, facilitant leur adoption clinique et répondant aux besoins spécifiques de la segmentation en oncologie
Passivity Filters for Bilateral Teleoperation with Variable Impedance Control
No full textInternational audienceIn robotic teleoperation, it is crucial to be able to dynamically adjust interactions with the environment. Drawing inspiration from human behavior during interactions, Variable Impedance Control (VIC) has been widely adopted to enhance robotic flexibility and adaptability. However, maintaining the passivity of such control systems remains a critical safety concern. This paper introduces an optimization-based framework for passive variable impedance control in bilateral teleoperation, combining the advantages of Passivity Filters (PF), Time-Domain Passivity (TDP) control, and Passive-Set-Position-Modulation (PSPM). The method solves an optimization problem aimed at dissipating the energy that could lead to a lack of passivity. The proposed method is assessed through experiments, illustrating its ability to keep the teleoperation system passive and safe under a variable impedance profile
Software design patterns for a STRIDE approach on an AUV fleet
No full textInternational audienceThe use of drones has become widespread in many fields, including critical ones. It means that these devices need to be protected against cyber attacks, and more generally, against contingencies, with a view to infrastructure resilience.This study simulates the operations of a drone fleet and assesses the impact of potential attacks or threats. It leverages the STRIDE threat modeling framework, which categorizes six types of threats: impersonation, tampering, repudiation, information disclosure, denial of service, and privilege escalation. Autonomous underwater vehicles (AUVs) are available, affordable, and can play a critical role for many applications; therefore, the systems we chose as analysis targets integrate AUVs and USVs as physical components, as well as digital avatars. This work is based on a model-based systems engineering (MBSE) approach. The use of models facilitates interoperability with third-party tools, enabling hybrid simulations and flexible integration of new devices, as well as global state capture via evaluators. Moreover, the use of well-established software design patterns ensures the modularity, reusability, and maintainability of the environment.Currently, our case study is to map an area of interest using a fleet of virtual drones, prior to deploying a mixed infrastructure. The process starts with data acquisition, followed by the refinement of data into information, then into knowledge, assuming the zone has been previously covered by a hydrographic study, and finally provides mission feedback. The potential risk lies in the manipulation of data or disruption of the drones' functionalities (such as movement, communication, etc.), which could compromise the mission's success and the integrity of the infrastructure.Future work will extend this framework to advanced threat and risk assessments (TARA), taking partial failures into account, exploring complex scenarios and drawing on more comprehensive metrics to propose appropriate security enhancement measures
Non-rigid motion compensation with skin deformation prediction for in situ bioprinting
No full textInternational audienceThis paper introduces a novel method of non-rigid motion compensation for in situ bioprinting. Most bioprinting platforms use open-loop systems, but it raises concerns about patient safety and suboptimal wound coverage in case of patient motion. To handle these issues, our method integrates an RGB-D camera to manage orientation and to predict deformations, along with a laser telemeter to regulate deposited material thickness. The proposed approach has been evaluated on a moving silicone platform that deforms at 0.8 Hz with a 4 mm in-plane amplitude and a 20 mm elevation amplitude. Our method resulted in a wound coverage error of less than 1 %. Comparative analysis demonstrates a 73.0% enhancement in deforming path following compared to existing methods. Additionally, by predicting surface motion, the method enables more precise control of layer height, with an error inferior to 0.1 mm.Cet article présente une nouvelle méthode de compensation des mouvements non rigides pour la bio-impression in situ. La plupart des plateformes de bio-impression utilisent des systèmes en boucle ouverte, mais cela pose des problèmes de sécurité pour le patient et de couverture sous-optimale de la plaie en cas de mouvement du patient. Pour résoudre ces problèmes, notre méthode intègre une caméra RGB-D pour gérer l'orientation et prédire les déformations, ainsi qu'un télémètre laser pour réguler l'épaisseur du matériau déposé. L'approche proposée a été évaluée sur une plateforme mobile en silicone qui se déforme à 0,8 Hz avec une amplitude de 4 mm dans le plan et une amplitude d'élévation de 20 mm. Notre méthode a permis d'obtenir une erreur de couverture de la plaie inférieure à 1 %. L'analyse comparative démontre une amélioration de 73,0 % dans le suivi de la trajectoire de déformation par rapport aux méthodes existantes. En outre, en prédisant le mouvement de la surface, la méthode permet un contrôle plus précis de la hauteur de la couche, avec une erreur inférieure à 0,1 mm
Unexpected microbial diversity in new Caledonia’s ultramafic ecosystems with conservation implications in a biodiversity hotspot
No full textData availability: DNA librairies: the Illumina MiSeq sequences are available under the following NCBI accession numbers: SAMN05786746 to SAMN05786777 for Rivière Blanche and Kopéto; SUB9939738, SUB9957890, SUB9966585, SUB9965709 and SUB9966666 for Goro; SAMN31953029 to SAMN31953058 and SAMN31953148 to SAMN31953177 for Maré; SRR30291378 to SRR30291403 for Bois du Sud and Tiébaghi.International audienceSoils harbour an incredible diversity of microorganisms that play crucial roles in ecosystem functioning. However, this biodiversity remains largely overlooked, with a poor understanding of how patterns form across landscapes. An eDNA metabarcoding approach was used to identify potential overarching patterns in fungal and bacterial communities from ultramafic ecosystems in New Caledonia, a renowned biodiversity hotspot. Our comprehensive analysis revealed several key findings, notably an important microbial diversity in the extreme environments of iron crust soils. Clear tendencies in phyla composition were also observed, with the fungal groups Ascomycota and Mucoromycota acting as potential indicators of land degradation (only in lateritic soils for Mucoromycota). For bacteria, Chloroflexi was characteristic of open vegetation, while Proteobacteria and Cyanobacteria were observed in higher relative abundances in the closed vegetation. The ectomycorrhizal fungal functional group was also found to be rich and unique, with a hypothetical endemism rate of 87%, and over-represented by the Cortinarius genus in rainforests and maquis (shrublands) dominated by ectomycorrhizal plants. Finally, each ultramafic Massif demonstrated a unique microbial community. Thus, our findings provide valuable insights into microbial ecology and emphasize the need for tailored conservation strategies for this biodiversity hotspot
Nginx a 20 ans : configuration avancée
No full textNational audienceDans l’article précédent, nous avions vu les principes généraux de configuration de Nginx, je vous avais promis de montrer comment réaliser un modèle de configuration permettant de déployer des services très facilement. C’est ce que je vous propose de voir dans les pages qui suivent
An Instruction Dataset for Extracting Quantum Cascade Laser Properties from Scientific Text Authors
No full textInternational audienceQuantum Cascade Lasers (QCL) are promising semiconductor lasers, compact and powerful, but of complex design. Availability of structured data of the QCL properties can support data mining activities that seek to understand the relationship between these properties, for instance between the design and performance features. The main open source of QCL data is in scientific text which in most cases is usually unstructured. One of the ways to extract and organize this data is by utilizing Information Extraction techniques. These techniques can accelerate the process of curating QCL properties data from scientific articles for further analysis. One of the main challenges in developing machine learning algorithms for extraction of QCL properties from text is lack of quality training data for these algorithms. Large Language Models (LLMs) have demonstrated great capabilities in materials property extraction from text. They however experience challenges with domain specific properties, for instance the heterostructure and design types in the QCL domain hence for adaptation. In this paper, we present an original instruction dataset for training and evaluation of large language models (LLMs) for QCL properties extraction from text. The data is generated by augmenting sample sentences from scientific articles with GPT-3.5 instruct with a few shot strategy. The dataset then is manually annotated with the help of QCL experts and is composed of 1300 rows of training examples consisting of an Instruction, Input Text and the Output
Uncovering the Intricacies and Synergies of Processor Microarchitecture Mechanisms using Explainable AI
No full textInternational audienceThis paper defines a data-driven methodology seamlessly combining machine learning (ML) and eXplainable Artificial Intelligence (XAI) techniques to address the challenge of understanding the intricate relationships between microarchitecture mechanisms with respect to system performance. By applying the SHapley Additive exPlanations (SHAP) XAI method, it analyzes the synergies of cache replacement, branch prediction, and hardware prefetching on instructions per cycle (IPC) scores. We validate our methodology by using the SPEC CPU 2006 and 2017 benchmark suites with the ChampSim simulator. We illustrate the benefits of the proposed methodology and discuss the major insights and limitations obtained from this study
A two-head loss function for deep Average-K classification
No full textInternational audienceAverage-K classification is an alternative to top-K classification in which the number of labels returned varies with the ambiguity of the input image but must average to K over all the samples. A simple method to solve this task is to threshold the softmax output of a model trained with the cross-entropy loss. This approach is theoretically proven to be asymptotically consistent, but it is not guaranteed to be optimal for a finite set of samples. In this paper, we propose a new loss function based on a multi-label classification head in addition to the classical softmax. This second head is trained using pseudo-labels generated by thresholding the softmax head while guaranteeing that K classes are returned on average. We show that this approach allows the model to better capture ambiguities between classes and, as a result, to return more consistent sets of possible classes. Experiments on two datasets from the literature demonstrate that our approach outperforms the softmax baseline, as well as several other loss functions more generally designed for weakly supervised multi-label classification. The gains are larger the higher the uncertainty, especially for classes with few samples
