Research Projects

In neonates and infants, the decision to operate on laryngotracheal stenosis is based on its clinical tolerance. However, there are intermediate forms, where the decision of major surgery is complex. Computational Fluid Dynamics can be used to model airflow in the pediatric airway. An efficient use of CFD requires both modeling a reliable 3D reconstruction of pediatric airways and solving the equations that govern the flow. These steps are time-consuming and sometimes not reproducible. PedIATrach aims at integrating computer vision approaches for automating the 3D models reconstruction and optimizing the reconstruction time by selecting the appropriate DICOM images.

Neurodevelopmental disorders (NDDs) include impairments in some domains of functioning, and their diagnoses are based on behavioral analysis with developmental acquisition lags. The analysis of mouse models of NDD, which is essential for identifying pathological mechanisms and improving the treatment of NDD, requires the identification of a wide range of complex and evolving behaviors to characterize pathological models and verify the efficiency of treatments. For this, we will use LMT, and to develop analytical tools for studying complex and evolving behaviors. Our objective is to analyze the behavior of 5 NDD mouse models at different ages.

ASDMOD aims at modelling behaviors of autistic (ASD) mouse models, caused by either genetic mutations, excessive gestational weight gain (GWG, detected today in half of women), or both. It focuses on a non-invasive, longitudinal, and blind behavioral study of male and female autistic mouse models (genetic and environmental models) compared with control mouse (same age, same sex). We thus propose a 3D behavioral analysis pipeline dedicated to the study of temporal series of single actions (individual, dyadic or group actions). This framework will rely on computer vision and deep learning techniques and will consider these actions time series at multiple time-scales.

Infective endocarditis (IE) is a life-threatening condition with high morbidity and mortality, difficult to diagnose, involving infection of the heart lining and valve destruction with vegetation formation. Early targeted antibiotic treatment can reduce the risk of fatal embolisms caused by vegetation detachment. Despite advances in diagnosis and treatment, IE's impact remains significant. We've developed a novel method using scanning electron microscopy for detailed vegetation analysis, which varies by bacterial species. To overcome current manual, subjective analysis, we aim to employ deep learning for automated, standardized detection and quantification of vegetations to enhance understanding and treatment.

The VRAK3D project integrates archaeology, architecture, and AI to enhance Cultural Heritage understanding. Focusing on the Glanum site, it has four phases: 1) Creating a 3D model integrating archaeological knowledge, aiding further discoveries. 2) Developing a formalized corpus of objects using ontologies like CRM-CIDOC, detailing morphology and inter-object relationships. 3) Utilizing deep learning to identify and localize objects of interest from photographs for 3D mapping. 4) Implementing VR for virtual site tours and knowledge interface. This project promises automated image and 3D model semantization, enriching the field with innovative technical applications.

The project employs 3D modeling for archaeological studies to automate stone-by-stone surveys and masonry analysis using deep learning and virtual reality. The goal is to streamline the process of architectural restitution by integrating data directly in 3D. Over two years, it will develop automatic detection of structural elements from extensive datasets of various archaeological sites, enhancing material identification and reducing the manual effort involved in traditional methods.

This project seeks to enhance medical imaging segmentation by aggregating diverse, annotated datasets and evaluating current tools to build a comprehensive national collection. It will develop an optimized, universally applicable segmentation solution through collaborative tool validation. Its novelty lies in: i) establishing and sharing enriched image databases to improve segmentation techniques, ii) integrating these methods with clinical workflows for real-time expert feedback, and iii) applying findings to physiological studies and nutritional intervention monitoring. The project fills gaps in existing databases, leverages innovative deep learning segmentation methods, and introduces a unique web tool for clinical application.

This service contract involves the photogrammetric survey of three underwater sites, which act as paradigms for scenarios where a robot's environment learning is essential for precise task execution. It includes the virtualization of surveys to create virtual sites for simulating robot behavior, aiding in training phases. Additionally, it establishes a Deep Reinforcement Learning framework within the UNITY simulation platform, enabling a transition to real-world experimentation, contingent only on mechanical constraints and available marine resources.

The MoDyPe3D project aims to enhance pathological understanding, surgical procedures, and patient care through customizable 3D modeling. By using multi-planar MRI acquisition, it visualizes pelvic organs in five planes, although manual segmentation is time-consuming (50 minutes per plane). We've successfully demonstrated 3D visualization of pelvic organ movement during exercises in women with prolapse. After creating a neural network database from our segmentations, we plan to automate MRI segmentation, significantly cutting time—a precursor to regular clinical use. This work progresses towards creating 'digital twins' for pre-operative planning.

ACTIVIS, an interdisciplinary project, aiming at improving scientific research on ASD by quantifying the atypical behaviors on which the diagnosis of autism is based. It gathers medical and scientific goals including diagnosis assistance, evaluation of therapeutic programs, and quantification of atypical behaviors. This quantification requires video analysis of autistic people. ACTIVIS addresses this problem by developing computer vision and learning approaches. Indeed, computer vision is an effective solution for studying the behavior of autistic children: It allows a non-intrusive and continuous capturing of behavioral patterns with low implementation costs and provides objective measure of the impact currently used cures and treatments.

DeepAI aims to assess the health of gorgonians in Marseille's Calanques National Park amidst global change, causing widespread invertebrate mortality. The project uses participatory science for data collection via scuba diving photography and AI for image analysis. It will develop a model for health assessment and sustainable monitoring through participatory science, involving a network of experts across disciplines. The initiative will enhance long-term biodiversity monitoring and provide data and insights to environmental managers, showcasing a model for a long-term observatory that merges human and analytical resources.

The exploitation of underwater resources and maritime security are of geostrategic importance, involving civilian, military, and scientific stakeholders. The use of autonomous or semi-autonomous vehicles, such as ROVs, AUVs, and ASVs, is growing for underwater inspection and security. The CAPRICA project aims to improve the identification of underwater targets using an AUV or ASV, despite challenges in precise navigation and low visibility. It proposes advanced sensors and signal processing to detect and characterize areas of interest. Defense and scientific applications, such as studying barracuda schools in collaboration with Septentrion Environment, are being explored.

Acquisition of 3D models of the seabed reconstructed by photogrammetry, labeling of these seabeds, and acquisition of side-scan sonar images of the seabed in these same areas. This project is a service contract (marché publique)

Pelvic-perineal static disorders are increasingly common with age, yet their pathophysiology remains poorly understood, complicating patient management. Our multidisciplinary project models pelvic organ dynamics, involving: Biomechanics: to simulate organ behavior under stress and create interactive mechanical models. MRI: to capture sequences showing organ responses during stress exertion. Imaging: to reconstruct 3D+t organ models and characterize observed deformations. 3D observations calibrate and validate these simulations, enhancing understanding of pathologies, surgical methods, and patient care. This customizable modeling is pivotal for advancing treatment strategies.

The MoDyPe project highlighted the need to characterize organ behavior using dynamic MRI for a better understanding of underlying phenomena; 2D MRI is routine, but 3D models are essential for deeper insights. The LIS and CRMBM have pioneered 3D studies on bladder movement via ultra-fast MRI. Building on this, the DynPelvis3D project collaborates with clinicians to extend observations to adjacent organs, focusing on: Testing MRI sequences for dynamic 3D organ observation Creating tools for visualizing and quantifying 3D organ deformations, aiding clinical visualization and numerical modeling. Furthermore, DynPelvis3D has spurred a clinical research proposal to the APHM, planning volunteer recruitment for MRI sequence development and patient inclusion in preliminary trials.

The project involves developing a new underwater vehicle and associated services, allowing for the monitoring, photographing, and modeling of natural organisms and human-made underwater installations. This project is led by a consortium formed around an intermediate size enterprise (ALSEAMAR), a small and medium-sized enterprise (COMEX), and three research laboratories (LSIS, I3S, and MAPIEM), all located in the Provence-Alpes-Côte d'Azur region.

The i-MareCulture project aims to promote European identity through underwater maritime heritage, which has traditionally been out of reach. It utilizes virtual tours, serious games, and augmented reality to make this heritage accessible. This interdisciplinary project is part of the European Digital Agenda and the H2020 program, aiming to connect education, research, and industry. It supports innovation and commercialization in the fields of virtual museums and digital heritage, aligning with the European strategy for a smart, sustainable, and inclusive economy.

The LORI project aims to develop a solution for object localization and recognition within a scene for the guidance of underwater vehicles. This initiative addresses the challenge of ensuring complete coverage of underwater sites during surveys to avoid incomplete 3D models and additional costs. Mobile submersibles or divers navigate targeted sites, capturing images processed through photogrammetry for realistic 3D reconstructions. The project tackles issues such as 'blind' scanning due to poor lighting by proposing real-time visual odometry for navigational guidance and clear indicators for image quality assessment. This results in a comprehensive hardware and software solution that streamlines the process of underwater documentation and minimizes time and resource expenditure.

Our interdisciplinary project connects underwater/naval archaeology with computer science, bringing together experts from France, Malta, and the USA. We aim to create an information system based on ontologies to manage digital content and expert knowledge in a uniform way. This system will integrate field data collected, in particular through underwater photogrammetry, a precise tool for underwater archaeology. We have used this technique, supported by ontologies, to automate the survey of a deep wreck in Malta. Subsequently, the project will extend to naval archaeology, in partnership with the University of Texas, to refine the structural knowledge of ships.

The Mediterranean, a "hotspot" of marine biodiversity, is facing an ecological crisis exacerbated by human activities, particularly in the coralligenous habitats, which are diverse but vulnerable. The PERFECT project adopts a multidisciplinary approach to conserve these ecosystems, focusing on the red coral, a key species. The project aims to establish conservation criteria and to study coral reproduction to improve techniques for population restoration. It incorporates underwater photogrammetry, ecology, genetics, and involves public awareness, scientific collaboration, and practical application in the management of marine areas.