Search

Learning inter-image similarity is crucial for 3D medical images self-supervised pre-training, due to their sharing of numerous same semantic regions. However, the lack of the semantic prior in metrics and the semantic-independent variation in 3D medical images make it challenging to get a reliable measurement for the inter-image similarity, hindering the learning of consistent representation for same semantics. We investigate the challenging problem of this task, i.e., learning a consistent representation between images for a clustering effect of same semantic features. We propose a novel visual similarity learning paradigm, Geometric Visual Similarity Learning, which embeds the prior of topological invariance into the measurement of the inter-image similarity for consistent representation of semantic regions. To drive this paradigm, we further construct a novel geometric matching head, the Z-matching head, to collaboratively learn the global and local similarity of semantic regions, guiding the efficient representation learning for different scale-level inter-image semantic features. Our experiments demonstrate that the pre-training with our learning of inter-image similarity yields more powerful inner-scene, inter-scene, and global-local transferring ability on four challenging 3D medical image tasks. Our codes and pre-trained models will be publicly available on https://github.com/YutingHe-list/GVSL., Comment: Accepted by CVPR 2023

The nature of thick-slice scanning causes severe inter-slice discontinuities of 3D medical images, and the vanilla 2D/3D convolutional neural networks (CNNs) fail to represent sparse inter-slice information and dense intra-slice information in a balanced way, leading to severe underfitting to inter-slice features (for vanilla 2D CNNs) and overfitting to noise from long-range slices (for vanilla 3D CNNs). In this work, a novel mesh network (MNet) is proposed to balance the spatial representation inter axes via learning. 1) Our MNet latently fuses plenty of representation processes by embedding multi-dimensional convolutions deeply into basic modules, making the selections of representation processes flexible, thus balancing representation for sparse inter-slice information and dense intra-slice information adaptively. 2) Our MNet latently fuses multi-dimensional features inside each basic module, simultaneously taking the advantages of 2D (high segmentation accuracy of the easily recognized regions in 2D view) and 3D (high smoothness of 3D organ contour) representations, thus obtaining more accurate modeling for target regions. Comprehensive experiments are performed on four public datasets (CT\&MR), the results consistently demonstrate the proposed MNet outperforms the other methods. The code and datasets are available at: https://github.com/zfdong-code/MNet, Accepted by IJCAI 2022

An effective backbone network is important to deep learning-based Deformable Medical Image Registration (DMIR), because it extracts and matches the features between two images to discover the mutual correspondence for fine registration. However, the existing deep networks focus on single image situation and are limited in registration task which is performed on paired images. Therefore, we advance a novel backbone network, XMorpher, for the effective corresponding feature representation in DMIR. 1) It proposes a novel full transformer architecture including dual parallel feature extraction networks which exchange information through cross attention, thus discovering multi-level semantic correspondence while extracting respective features gradually for final effective registration. 2) It advances the Cross Attention Transformer (CAT) blocks to establish the attention mechanism between images which is able to find the correspondence automatically and prompts the features to fuse efficiently in the network. 3) It constrains the attention computation between base windows and searching windows with different sizes, and thus focuses on the local transformation of deformable registration and enhances the computing efficiency at the same time. Without any bells and whistles, our XMorpher gives Voxelmorph 2.8% improvement on DSC , demonstrating its effective representation of the features from the paired images in DMIR. We believe that our XMorpher has great application potential in more paired medical images. Our XMorpher is open on https://github.com/Solemoon/XMorpher, Comment: accepted by MICCAI 2022

International audience; Rib fracture is a common disease that requires prompt treatment. This study focuses on developing a rib fracture diagnosis deep learning method using contralateral, contextual, and edge enhanced modules and evaluating its detection performance. A novel rib fracture diagnosis method was designed, named CCE-Net. To evaluate the performance of this method, 1639 digital radiography (DR) images were enrolled. Fracture features were extracted for three modules: contralateral, contextual, and edge enhanced modules. These modules can be used to identify fracture features in rib DR images, imitating the experience of broad-certificated radiologists. The contralateral module assists in diagnosing rib fractures by comparing the difference between the detected target region and the contralateral region. The contextual module helps to aid rib fracture detection by extracting contextual features. The edge enhanced module improves the accuracy of fracture detection by enhancing the edge information of the rib bone. The head of this two-stage detection network uses the multi-path fusion mechanism as the main architecture to integrate and utilize the above modules. The qualitative results show that with the ground truth of rib fracture as the evaluation standard, CCE-Net can achieve a better visual effect of fracture detection than other methods. The quantitative results show that CCE-Net can achieve the best performance in various detection indicators include AP50 0.911, AP75 0.794, AP25 0.913, and Recall 0.934. Experimental results show that CCE-Net can acquire the excellent ability of rib fracture diagnosis. We invasion that this approach will be applied to clinical study. © 2022 Elsevier Ltd

International audience

International audience

NUMÉRO SPÉCIAL ANR TECSAN - Technologies pour la santé et l'autonomie; International audience; n.a

International audience; Beyond the advances made in computer-assisted interventions and robotic systems, the demand for more efficient and safer therapies remains challenging. Thus, if it is possible to improve the instrument tracking, steering, and target localization, to miniaturize the sensors and actuators, and to conduct preoperatively planned minimally invasive therapies, we still need new resources to achieve permanent destruction of abnormal tissues or suppression of pathological processes. Most of the physics-based (or energy-based) therapeutic principles at our disposal have been established a long time ago, but their actions on basic cellular and molecular mechanisms are not yet fully understood. They all have a wide spectrum of clinical targets in terms of organs and pathologies, modes of application (external, interstitial, intraluminal, etc.) with advantages and side-effect drawbacks, proven indications, and contraindications. Some of them may still face controversies regarding their outcomes. This short article, mainly focused on tumor destruction, briefly reviews in its first part some of these techniques and sketches the next generation under investigation. The former include radio frequency (RF), high-intensity focused ultrasound (HiFU), microwaves, and cryotherapy, of which all are temperature based. Laser-based approaches [e.g., photodynamic therapy (PDT) at large] are also discussed. Radiotherapy and its variants (hadrontherapy, brachytherapy, Gamma Knife, and CyberKnife) remain, of course, as the reference technique in cancer treatment. The next breakthroughs are examined in the second part of the article. They are based on the close association between imaging agents, drugs, and some stimulation techniques. The ongoing research efforts in that direction show that, if they are still far from clinical applications, strong expectations are made. From the point of view of interventional planning and image guidance, all of them share a lot of concerns.

International audience

International audience; Medical robotics, computer- assisted surgery (CAS), image-guided therapy (IGT), and the like emerged more than 20 years ago, and many advances have been made since. Conferences and workshops have been organized; scientific contributions, position papers, and patents have been published; new academic societies have been launched; and companies were created all over the world to propose methods, devices, and systems in the area. Researchers in robotics, computer vision and graphics, electronics, mechanics, biomedical engineers, physicians, and surgeons have been involved, thus demonstrating the enthusiasm for this emerging field. Their commitments emphasize the transdisciplinary nature of the efforts to be made. However, the effective dissemination of CAS-IGT systems in medical disciplines remains limited. There are several reasons that may explain this situation, among which the effective demonstration of patient benefits and cost savings, the reluctance of surgeons or therapists to use them, and, of course, the technological breakthroughs are still expected. This series of articles attempts to point out some of them and will also show that many opportunities in computer-assisted interventions are open in the near future.

International audience; Moment-based methods cover a wide spectrum of applications and deserve to be confronted to other attractive and competitive methods. The aim of this article is to identify the main contributions of the moments rather than entering into the methodological details and examples.

International audience; Moment functions have been defined in [1] and important properties such as invariance and robustness to noise have been reviewed in the second paper [2]. Before addressing applications of moments, another feature has to be discussed, the computational load. The complexity of image analysis methods, in other words the number of operations they require to achieve a given task, iteratively or not, may lead to practical limitations when dealing with large data sets (2D or 3D image sequences) and time constraints. This issue is also of concern for moments in particular when high orders have to be computed. Special attention must therefore be paid to fast computation. The continuous-to-discrete transform may also affect the analytical properties we must preserve (i.e. invariance, orthogonality, etc.) by introducing numerical errors. The problem of accurate computation of moments should thus be addressed. These two aspects are examined in this third paper.

International audience; The several moment families have been reviewed in a first paper [1]. A classification was proposed in order to get a better understanding of their relations. More attention was given to orthogonal moments (in particular Legendre, Zernike, Tchebichef, Krawtchouk, Racah, dual Hahn). Important properties for computer vision applications were just sketched, among which invariance and robustness to noise. These properties may drive the choice of moments when addressing a specific problem. A short, thus non-exhaustive, review of the literature on these issues is proposed in this second paper.

Although recent enthousiasm has emerged for Systems Biology, it is of major importance to identify the roots it has with computational (mathematical) modeling. In fact, major contributions have been made for decades with the aim to quantitatively analyze and model the function of living systems in order, ultimately, to better understand the underlying constituents and collective behaviors and use them for diagnosis and therapeutic purposes. However, the impressive evolution of technological resources and methods allows today to revisit these early attempts and to bring to light new concepts, targets, and expectations. This chapter, after a tentative definition of the generic elements behind Systems Biology and Physiology, will provide a review of current efforts devoted to multimodal, multilevel, multiresolution approaches, all being addressed from the joint observational, modeling and information processing points of view. The several theoretical frames at our disposal will also be addressed, in particular the capability to handle multiple modeling for- malisms. Examples from the literature and the research conducted by the authors will exemplify these multidisciplinary developments and results. The forthcoming challenges to be faced will then be outlined.

International audience; This first paper was aimed at providing the basic formulations of moments, a classification and an introductory bibliography. The moment-based approaches have a number of interesting features. They have a wide range of orthogonal and non-orthogonal basis functions and are simple to compute whatever the order required. The image sampling can be either rectangular or polar, based on uniform or non-uniform lattices. Their optimal choice to deal with a given problem is however not obvious according to the requirements to face. There are of course many other issues to address. Among the important properties to consider there are (i) the invariance to scale, translation, and orientation, etc. (ii) the robustness to degradations, noise, to changing conditions (illumination) or blurring and (iii) to object variations (multiple appearances, occlusions and deformations). Moment computations have also a cost that is sometimes considered too high for certain applications: this explains that a special attention has been devoted to acceleration techniques and VLSI implementations. These important aspects for computer vision at large and the most meaningful works in medical imaging will be examined in the next papers.

Biological and medical imaging technologies have made impressive advances over the last decades. Some of their dimensions are health care driven (from screening, diagnosis assessment to therapy), others technology driven (bio-arrays, micro- to macro imaging) or biology driven (genes to proteins and cells). Computational modelling and information technology in biology and medicine is also becoming to have a major role in the interdisciplinary attempt to elucidate functions of living systems and structure-to function inter-relations. However, the recursion between description levels, at all space and time scales, for highly dynamical, interacting and complex systems, still requires new breakthroughs in observational tools closely coupled to sound information processing for confrontation to theoretical models.

This third and last paper in this series is devoted to biological imaging, e.g nano and micro imaging. Although a number of concerns are shared when going to these scales, the conditions of observation, the objects under study and the problems to address make this research very challenging by the cautions that must be taken to control the experiments, a mandatory condition to get conclusive results and to infer sound conclusions.

International audience; no abstract

International audience; .

http://www.roczniki-amb.com/roczniki_2005/ http://www.roczniki-amb.com/roczniki_2005/volumes/18_Coatriux_JL_A%20grand%20challenge.pdf; Computational modelling, nano-bioscience and information technology in biology and medicine will play a major role in the interdisciplinary attempts to elucidate structures and functions of living systems. Developing tools capable to integrate the new advances and make benefit of them is crucial: accumulation of data and knowledge base with only storage and retrieval capabilities will have a poor impact if they are not made "active" or "operational". This is where models will play a central role in offering, not only sound ways for representation or simulation, but also the appropriate frames to put the players in the right place, with intra- and inter-level coupling and multisource handling. This paper advocated that sequential observations of multiple and complex mechanisms will be of limited interest to understand the inter-relations that are occurring at the same time, and therefore, that designing multimodal, multilevel and multiscale experiments, matched with these models, are of major importance.

Coupling computational modeling and information processing in biology and medicine is a major challenge for better comprehending structures and functions of living systems. Signal processing should extract the relevant information required to explore complex organization levels, at all space and time scales. Advances coming from applied physics and mathematics are challenged by extremely hot topics in biology and medicine. The biomedical scene has proven to be the most difficult to address due to the fact that biomedical processes involve nonGaussian, nonlinear, and nonstationary components. This paper provides some clues on processing schemes such as time and frequency transforms, blind signal separation, independent component analysis, empirical mode decomposition, particle methods and Kernel methods that may help in lessening the ambiguity about the observed components of the mixtures to be handled and, this way, facilitating their matching with models.

International audience

Learning how the human heart develops is essential in cardiology education. It lets us understand a normal heart's morphology and also explains the formation of congenital heart defects. However, learning cardiac embryology is difficult because the processes underlying the heart formation are complex. This complexity is caused by the fact that complicated organs like the heart are generated from only a couple of cells. The overall objective of the project presented was to design and develop a 3D model capable of rendering the dynamics of the various stages of normal heart development, beginning at fertilization until its final form.

Signal processing offers a wide spectrum of theories, methods, and algorithms for addressing a variety of problems ranging from noise reduction, restoration, detection (of events or changes), spatiotemporal dynamics estimation, source localization, and pattern recognition. However, the classical assumptions (stationarity, linearity, etc.) usually do not apply in real situations. Recent advances, such as time-scale and time-frequency transforms, data fusion, long-range dependence, and higher order moments, do not always provide sufficiently robust solutions. In this article, the basic properties and generic features of biomedical signals are examined using a wide range of examples. Algorithmic results are presented to show not only the potential performance but also the limitations of the processing resources at our disposal. The last section describes and discusses signal matching, scenario recognition, and data fusion.

National audience; Monnet : une application du travail coopératif en imagerie médicale

Some concepts that seem fundamental for the quality evaluation of data visualization in general and of medical data in particular are presented., Apresentam-se neste trabalho alguns conceitosque parecem fundamentais no problema da avaliação daqualidade da visualização em geral e de dados médicos emparticular.

International audience; Le travail présenté dans cet article s'inscrit dans une démarche globale visant l'interprétation de la cinétique cardiaque après analyse des vaisseaux du myocarde. Il est focalisé ici sur le problème d'extraction de caractéristiques géométriques intermédiaires dont l'évolution temporelle doit servir de base à la construction des informations symboliques nécessaires à la description locale et globale du mouvement. Les paramètres géométriques choisis sont liés à la définition de point virtuel de référence, de courbure, de torsion et d'axes d'inertie. Les résultats obtenus permettent un aperçu du mouvement, l'analyse détaillée des déformations locales et l'obtention de symboles décrivant la cinétique de l'organe par parties homogènes. L'exposé est essentiellement axé sur les méthodes d'extraction de paramètres.

International audience; In computational vision research, implicit or explicit knowledge is used for a number of tasks such as edge or line detection, stereo matching, motion tracking and three-dimensional representation. This knowledge deals with the image capture, the objects, the scenes under study as well as the set of processing algorithms and can be classified along several lines (intrinsic vs extrinsic features, time-space dimensions, low vs high level, generic vs specific, local vs global, etc.). These multiple points of view must be recognized as "experiential" or "in depth" modeling aimed at ordering the procedures, coupling data and knowledge, controlling potential failures and deviations, or in other words, at solving a given problem. These issues are illustrated through the 3-D reconstruction of the cardiac vascular network from a pair of angiographic image sequence. This highly underdetermined problem obliges to incorporate and distribute a priori knowledge in all the steps involved in the reconstruction process and beyond. They cover : the image acquisition (choice of incidences); the extraction of vessels (contour lines, centerlines); complex configuration analysis (crossings, superpositions); 2-D vessel motions (disambiguation from the tracking); 3-D static reconstruction (model based approach); 3-D sequence reconstruction (3-D motion estimation); morphological representation (image rendering); functional interpretation (kinetic properties).

The purpose of this study was to investigate the effect of pressure losses during the cardiac cycle on the cross-sectional shape of a partially compliant coronary artery stenosis. A mathematical model of pulsatile flow through a partially collapsible stenosis was used to calculate changes in minimum stenosis diameter in response to changes in intraluminal pressure. Nine stenoses were simulated, ranging from 50 to 90% area reduction with different fractions of flexible wall circumference (1/4, 1/3, and 1/2). For typical flow rates during maximum vasodilation the minimum diameter of simulated severe stenoses decreased by over 50% from systole to diastole, thereby substantially altering the cross-sectional area of the stenosis. These results may explain some of the discrepancies that have been reported from pathologic and angiographie studies of coronary artery stenosis morphology.

A moment-based operator for 3D edge detection has been already proposed for isotropic description of medical data sets. This paper presents its generalization to anisotropic volume data. Two solutions, based on implicit "deformation" and "interpolation" model, are briefly described. This comparison is conducted on real data and discussed.

A QRS detection algorithm has been developed for use in body surface mapping of ventricular tachycardia. To detect the QRS complexes the algorithm uses three computed signals: the sum of the absolute value of all leads, the sum of the absolute value of the differences of consecutive samples in all leads and, the second order derivative of the first mentioned.

To evaluate a newly designed Doppler ultrasound transducer for registration of turbulence, simultaneous point blood velocity measurements using hot-film anemometry were performed downstream of smooth graded stenoses on the abdominal aorta in pigs. Signal analysis included ensemble averaging, turbulence analysis and visualization of one-dimensional velocity profiles. We conclude, that pulsed Doppler ultrasound as compared to hot-film anemometry is able to register one-dimensional velocity profiles and estimate turbulent stresses in the poststenotic flow field.