Search

This thesis presented conditional-random-field-based approaches for medical applications on diverse omics data. This methodology allowed leveraging more complex, structural information and notable assets from graph theory, particularly interesting to express intricate biological properties. We demonstrated their usefulness for clustering and feature selection towards classification. Their relevance was exemplified over several medical applications and omics data.First, we proposed a generic and resilient feature selection and classification pipeline we developed for COVID-19 patients staging and outcome prediction using only CT scans and clinical information. Relying on an automated segmentation technique, we extracted imaging information. After a required step of dimensionality reduction, we singled out few relevant factors for classification. We obtained promising performance outperforming radiologist experts on all the tasks. We further extended and adapted our methodology to cope with other different omics data, diseases, and medical expectations.Second, we focused on a clustering process towards the determination of a clinically relevant gene signature for pan-cancer lesions characterization. Oncology is a perfectly suited area for this kind of approach as tumors present a high heterogeneity while being a major affliction worldwide. Many studies are involved in its description through genomics. However, the task's complexity dwells in the data's large dimensionality and the experimental cost for identifying unknown gene functions. We highlighted our compact signature's relevance by resorting to unsupervised and supervised tumor types and subtypes distinction combined with statistically significant biological considerations.Finally, we formulated a new higher-order distance learning framework for feature selection and weighting, relying on conditional random fields and clustering. We proposed a mathematical optimization method for its resolution able to handle the high-order information complexity efficiently. Strong from this paradigm's expressiveness, we investigated the use of high-order graph theory properties as cliques, eccentricity, connectivity, or path lengths. We established those attributes' informativeness in classification settings and reported superior results than with standard approaches., Cette thèse introduit l’usage d’approches reposant sur les « conditional random fields » à diverses applications médicales et données omiques. Ces méthodes permettent de tirer parti au mieux d’informations structurelles lourdes à interpréter et analyser. En particulier, l’emploi de la théorie des graphes d’ordre supérieur revêt un intérêt majeur pour l’expression des relations biologiques complexes. Nous démontrons leur pertinence dans les domaines du « clustering » et de la sélection de variables pour la classification. Nous nous sommes appuyés sur plusieurs applications médicales et données omiques pour mettre ces résultats en lumière. Dans un premier temps, nous avons proposé un système générique et résilient de sélection de variables et de classification que nous avons développé pour déterminer la sévérité de la maladie de patients atteints de la COVID-19. Dans ce but, nous nous sommes appuyés sur des informations extraites de segmentations automatiques des organes et zones lésées que nous avons combinées avec des informations cliniques. Nous avons identifé un nombre restraint de facteurs déterminants la classification. Nous avons obtenu des performances prometteuses dépassant celles de radiologues experts sur les tâches considérées. Nous avons étendu plus avant et adapté cette méthodologie pour traiter d'autres données omiques, maladies et attendus médicaux. Par la suite, nous avons étudié un procédé de clustering pour la définition d'une signature de gènes présentant un intérêt clinique vis-à-vis de la caractérisation pan-cancer de lésions. Bien des études se sont essayées à la description du cancer grâce à la génomique. Cependant, la grande dimensionalité des données représente un formidable obstacle. Nous avons prouvé la pertinence de la signature génétique très compacte générée par notre méthode en recourrant à des approches supervisées et non-supervisées pour la caractérisation des types et sous-types de tumeurs. Finalement, nous avons défini une nouvelle approche d'apprentissage de distance d'ordre supérieur à visée de sélection et de pondération de variables. Fort de la grande expressivité de ce paradigme, nous avons exploré diverses propriétés de théorie des graphes d'ordre supérieur et avons établi que, dans le cadre d'une tâche de classification, ils possèdent une grande expressivité et permettent d'obtenir des résultats supérieurs à ceux des méthodes standards.

International audience; Finding the global minimum of a nonconvex optimization problem is a notoriously hard task appearing in numerous applications, from signal processing to machine learning. Simulated annealing (SA) is a family of stochastic optimization methods where an artificial temperature controls the exploration of the search space while preserving convergence to the global minima. SA is efficient, easy to implement, and theoretically sound, but suffers from a slow convergence rate. The purpose of this work is twofold. First, we provide a comprehensive overview on SA and its accelerated variants. Second, we propose a novel SA scheme called curious simulated annealing, combining the assets of two recent acceleration strategies. Theoretical guarantees of this algorithm are provided. Its performance with respect to existing methods is illustrated on practical examples.

International audience; Detecting change through multi-image, multi-date remote sensing is essential to developing and understanding of global conditions. Despite recent advancements in remote sensing realized through deep learning, novel methods for accurate multi-image change detection remain unrealized. Recently, several promising methods have been proposed to address this topic, but a paucity of publicly available data limits the methods that can be assessed. In particular, there exists limited work on categorizing the nature and status of change across an observation period. This paper introduces the first labeled dataset available for such a task. We present an open-source change detection dataset, termed QFabric, with 450,000 change polygons annotated across 504 locations in 100 different cities covering a wide range of geographies and urban fabrics. QFabric is a temporal multi-task dataset with 6 change types and 9 change status classes. The geography and environment metadata around each polygon provides context that can be leveraged to build robust deep neural networks. We apply multiple benchmarks on our dataset for change detection, change type and status classification tasks. Project page: https://sagarverma.github.io/qfabric Land Cleared Construction Midway Construction Midway Construction Done Operational Commercial Industrial.

International audience; Unlike their biological counterparts, simple artificial neural networks are unable to retain information from their past state to influence their behavior. In this contribution, we propose to consider new nonlinear activation functions, whose outputs depend both from the current and past inputs through a hysteresis effect. This hysteresis model is developed in the framework of convolutional neural networks. We then show that, by choosing the nonlinearity in the vast class of rational functions, the identification of the weights amounts to solving a rational optimization problem. For the latter, recent methods are applicable that come with global optimality guarantee, contrary to most optimization methods used in the neural network community. Finally, simulations show that such hysteresis nonlinear activation functions cannot be approximated by traditional ones and illustrate the effectiveness of our weight identification method.

International audience; This paper presents a quantitative approach to demonstrate the robustness of neural networks for tabular data. These data form the backbone of the data structures found in most industrial applications. We analyse the effect of various widely used techniques we encounter in neural network practice, such as regularization of weights, addition of noise to the data, and positivity constraints. This analysis is performed by using three state-of-the-art techniques, which provide mathematical proofs of robustness in terms of Lipschitz constant for feed-forward networks. The experiments are carried out on two prediction tasks and one classification task. Our work brings insights into building robust neural network architectures for safety critical systems that require certification or approval from a competent authority.

Based on its great successes in inference and denosing tasks, Dictionary Learning (DL) and its related sparse optimization formulations have garnered a lot of research interest. While most solutions have focused on single layer dictionaries, the recently improved Deep DL methods have also fallen short on a number of issues. We hence propose a novel Deep DL approach where each DL layer can be formulated and solved as a combination of one linear layer and a Recurrent Neural Network, where the RNN is flexibly regraded as a layer-associated learned metric. Our proposed work unveils new insights between the Neural Networks and Deep DL, and provides a novel, efficient and competitive approach to jointly learn the deep transforms and metrics. Extensive experiments are carried out to demonstrate that the proposed method can not only outperform existing Deep DL, but also state-of-the-art generic Convolutional Neural Networks., Accepted by ICASSP 2021. arXiv admin note: substantial text overlap with arXiv:2002.07898

International audience; Deep learning has achieved great success in processing large size medical images such as histopathology slides. However, conventional deep learning methods cannot handle the enormous image sizes; instead, they split the image into patches which are exhaustively processed, usually through multi-instance learning approaches. Moreover and especially in histopathology, determining the most appropriate magnification to generate these patches is also exhaustive: a model needs to traverse all the possible magnifications to select the optimal one. These limitations make the application of deep learning on large medical images and in particular histopathological images markedly inefficient. To tackle these problems, we propose a novel spatial and magnification based attention sampling strategy. First, we use a down-sampled large size image to estimate an attention map that represents a spatial probability distribution of informative patches at different magnifications. Then a small number of patches are cropped from the large size medical image at certain magnifications based on the obtained attention. The final label of the large size image is predicted solely by these patches using an end-to-end training strategy. Our experiments on two different histopathology datasets, the publicly available BACH and a subset of the TCGA-PRAD dataset, demonstrate that the proposed method runs 2.5 times faster with automatic magnification selection in training and at least 1.6 times faster than using all patches in inference as the most of state-of-the-art methods do, without loosing in performance.

Simple Summary As a rapid-development research field, radiomics-based analysis has been applied to many clinical problems. However, the reproducibility of the radiomics studies remain challenging especially when data suffers from scanner effects, a kind of non-biological variations introduced by different image acquiring settings. This study aims to investigate how the image preprocessing methods (N4 bias field correction and image resampling) and the harmonization methods (intensity normalization methods working on images and ComBat method working on radiomic features) help to remove the scanner effects and improve the radiomics reproducibility in brain MRI radiomics. Abstract In brain MRI radiomics studies, the non-biological variations introduced by different image acquisition settings, namely scanner effects, affect the reliability and reproducibility of the radiomics results. This paper assesses how the preprocessing methods (including N4 bias field correction and image resampling) and the harmonization methods (either the six intensity normalization methods working on brain MRI images or the ComBat method working on radiomic features) help to remove the scanner effects and improve the radiomic feature reproducibility in brain MRI radiomics. The analyses were based on in vitro datasets (homogeneous and heterogeneous phantom data) and in vivo datasets (brain MRI images collected from healthy volunteers and clinical patients with brain tumors). The results show that the ComBat method is essential and vital to remove scanner effects in brain MRI radiomic studies. Moreover, the intensity normalization methods, while not able to remove scanner effects at the radiomic feature level, still yield more comparable MRI images and improve the robustness of the harmonized features to the choice among ComBat implementations.

ObjectivesLumacaftor-ivacaftor is a cystic fibrosis transmembrane conductance regulator (CFTR) modulator known to improve clinical status in people with cystic fibrosis (CF). This study aimed to assess lung structural changes after one year of lumacaftor-ivacaftor treatment, and to use unsupervised machine learning to identify morphological phenotypes of lung disease that are associated with response to lumacaftor-ivacaftor.MethodsAdolescents and adults with CF from the French multicenter real-world prospective observational study evaluating the first year of treatment with lumacaftor-ivacaftor were included if they had pretherapeutic and follow-up chest computed tomography (CT)-scans available. CT scans were visually scored using a modified Bhalla score. A k-mean clustering method was performed based on 120 radiomics features extracted from unenhanced pretherapeutic chest CT scans.ResultsA total of 283 patients were included. The Bhalla score significantly decreased after 1 year of lumacaftor-ivacaftor (−1.40±1.53 points compared with pretherapeutic CT; p1) ≥5 under lumacaftor–ivacaftor than those in the other clusters (54% of responders versus 32% and 33%; p=0.01).ConclusionOne year treatment with lumacaftor-ivacaftor was associated with a significant visual improvement of bronchial disease on chest CT. Radiomics features on pretherapeutic CT scan may help in predicting lung function response under lumacaftor-ivacaftor.

The vast majority of semantic segmentation approaches rely on pixel-level annotations that are tedious and time consuming to obtain and suffer from significant inter and intra-expert variability. To address these issues, recent approaches have leveraged categorical annotations at the slide-level, that in general suffer from robustness and generalization. In this paper, we propose a novel weakly supervised multi-instance learning approach that deciphers quantitative slide-level annotations which are fast to obtain and regularly present in clinical routine. The extreme potentials of the proposed approach are demonstrated for tumor segmentation of solid cancer subtypes. The proposed approach achieves superior performance in out-of-distribution, out-of-location, and out-of-domain testing sets.

Patient-specific models of blood flow are being used clinically to diagnose and plan treatment for coronary artery disease. A remaining challenge is bridging scales from flow in arteries to the micro-circulation supplying the myocardium. Previously proposed models are descriptive rather than predictive and have not been applied to human data. The goal here is to develop a multiscale patient-specific model enabling blood flow simulation from large coronary arteries to myocardial tissue. Patient vasculatures are segmented from coronary computed tomography angiography data and extended from the image-based model down to the arteriole level using a space-filling forest of synthetic trees. Blood flow is modeled by coupling a 1D model of the coronary arteries to a single-compartment Darcy myocardium model. Simulated results on five patients with non-obstructive coronary artery disease compare overall well to [$$^{15}$$ 15 O]$$\text {H}_{{2}}$$ H 2 O PET exam data for both resting and hyperemic conditions. Results on a patient with severe obstructive disease link coronary artery narrowing with impaired myocardial blood flow, demonstrating the model’s ability to predict myocardial regions with perfusion deficit. This is the first report of a computational model for simulating blood flow from the epicardial coronary arteries to the left ventricle myocardium applied to and validated on human data.

This work is concerned with evolution equations and their forward-backward discretizations, and aims at building bridges between differential equations and variational analysis. Our first contribution is an estimation for the distance between iterates of sequences generated by forward-backward schemes, useful in the convergence and robustness analysis of iterative algorithms of widespread use in numerical optimization and variational inequalities. Our second contribution is the approximation, on a bounded time frame, of the solutions of evolution equations governed by accretive (monotone) operators with an additive structure, by trajectories constructed by interpolating forward-backward sequences. This provides a short, simple and self-contained proof of existence and regularity for such solutions; unifies and extends a number of classical results; and offers a guide for the development of numerical methods. Finally, our third contribution is a mathematical methodology that allows us to deduce the behavior, as the number of iterations tends to \begin{document}$ +\infty $\end{document} , of sequences generated by forward-backward algorithms, based solely on the knowledge of the behavior, as time goes to \begin{document}$ +\infty $\end{document} , of the solutions of differential inclusions, and viceversa.

International audience; Contrast-enhanced medical images offer vital insights for the accurate diagnosis, characterization and treatment of tumors, and are routinely used worldwide. Acquiring such images requires to inject the patient intravenously with a gadolinium-based contrast agent (GBCA). Although GBCAs are considered safe, recent concerns about their accumulation in the body tilted the medical consensus towards a more parsimonious usage. Focusing on the case of brain magnetic resonance imaging, this paper proposes a deep learning method that synthesizes virtual contrast-enhanced T1 images as if they had been acquired after the injection of a standard 0.100 mmol/kg dose of GBCA, taking as inputs complementary imaging modalities obtained either after a reduced injection at 0.025 mmol/kg or without any GBCA involved. The method achieves a competitive structural similarity index of 94.2%. Its asymptotic performance is estimated, and the most important input modalities are identified.

International audience; Image segmentation is one of the most popular problems in medical image analysis. Recently, with the success of deep neural networks, these powerful methods provide state of the art performance on various segmentation tasks. However, one of the main challenges relies on the high number of annotations that they need to be trained, which is crucial in medical applications. In this paper, we propose an unsupervised method based on deep learning for the segmentation of kidney grafts. Our method is composed of two different stages, the detection of the area of interest and the segmentation model that is able, through an iterative process, to provide accurate kidney draft segmentation without the need for annotations. The proposed framework works in the 3D space to explore all the available information and extract meaningful representations from Dynamic Contrast-Enhanced and T2 MRI sequences. Our method reports a dice of 89.8 ± 3.1%, Hausdorff distance at percentile 95% of 5.8±0.41mm and percentage of kidney volume difference of 5.9 ± 5.7% on a test dataset of 29 patients subject to a kidney transplant.

International audience; The year 2020 witnessed a heavy death toll due to COVID-19, calling for a global emergency. The continuous ongoing research and clinical trials paved the way for vaccines. But, the vaccine efficacy in the long run is still questionable due to the mutating coronavirus, which makes drug re-positioning a reasonable alternative. COVID-19 has hence fast-paced drug re-positioning for the treatment of COVID-19 and its symptoms. This work builds computational models using matrix completion techniques to predict drug-virus association for drug re-positioning. The aim is to assist clinicians with a tool for selecting prospective antiviral treatments. Since the virus is known to mutate fast, the tool is likely to help clinicians in selecting the right set of antivirals for the mutated isolate. The main contribution of this work is a manually curated database publicly shared, comprising of existing associations between viruses and their corresponding antivirals. The database gathers similarity information using the chemical structure of drugs and the genomic structure of viruses. Along with this database, we make available a set of state-of-the-art computational drug re-positioning tools based on matrix completion. The tools are first analysed on a standard set of experimental protocols for drug target interactions. The best performing ones are applied for the task of re-positioning antivirals for COVID-19. These tools select six drugs out of which four are currently under various stages of trial, namely Remdesivir (as a cure), Ribavarin (in combination with others for cure), Umifenovir (as a prophylactic and cure) and Sofosbuvir (as a cure). Another unanimous prediction is Tenofovir alafenamide, which is a novel Tenofovir prodrug developed in order to improve renal safety when compared to its original counterpart (older version) Tenofovir disoproxil. Both are under trail, the former as a cure and the latter as a prophylactic. These results establish that the computational methods are in sync with the state-of-practice. We also demonstrate how the drugs to be used against the virus would vary as SARS-Cov-2 mutates over time by predicting the drugs for the mutated strains, suggesting the importance of such a tool in drug prediction. We believe this work would open up possibilities for applying machine learning models to clinical research for drug-virus association prediction and other similar biological problems.

International audience; Rationale and Objectives: The purpose of this study was to develop and validate an algorithm that can automatically estimate the amount of coronary artery calcium (CAC) from non-enhanced ECG-gated computed tomography (CT) cardiac volume acquisitions by using multiple convolutional neural networks (CNN).Materials and Methods: Our method used an ensemble of 5 CNN with 3D U-Net architecture trained on a database of 783 CT examinations to detect and segment coronary artery calcifications in a 3D volume. The Agatston score, the conventional CAC scoring, was then computed slice by slice from the resulting segmentation mask and compared to the ground truth manually estimated by radiologists.The quality of the estimation was assessed with the Concordance Index (C-index) on a separate testing set of 98 independent CT examinations.Results: The final ensemble model yields a C-index of 0.9509 951 on the testing set. The remaining errors of the method were mainly observed on small-size and/or low-density calcifications, or calcifications located near the mitral valve or ring. Our model was also compared to other methods from the literature showing state-of-the-art performance for CAC quantification.Conclusion: The deep learning-based method we proposed to compute automatically the CAC score from a non-enhanced ECG-gated cardiac CT is fast, robust and could improve workflow efficiency, eliminating the time spent on manually selecting coronary calcifications to compute the Agatston score.

Over the last decade, the combination of compressed sensing (CS) with acquisition over multiple receiver coils in magnetic resonance imaging (MRI) has allowed the emergence of faster scans while maintaining a good signal-to-noise ratio (SNR). Self-calibrating techniques, such as ESPiRIT, have become the standard approach to estimating the coil sensitivity maps prior to the reconstruction stage. In this work, we proceed differently and introduce a new calibration-less multi-coil CS reconstruction method. Calibration-less techniques no longer require the prior extraction of sensitivity maps to perform multi-coil image reconstruction but usually alternate estimation sensitivity map estimation and image reconstruction. Here, to get rid of the nonconvexity of the latter approach we reconstruct as many MR images as the number of coils. To compensate for the ill-posedness of this inverse problem, we leverage structured sparsity of the multi-coil images in a wavelet transform domain while adapting to variations in SNR across coils owing to the OSCAR (octagonal shrinkage and clustering algorithm for regression) regularization. Coil-specific complex-valued MR images are thus obtained by minimizing a convex but nonsmooth objective function using the proximal primal-dual Condat-Vù algorithm. Comparison and validation on retrospective Cartesian and non-Cartesian studies based on the Brain fastMRI data set demonstrate that the proposed reconstruction method outperforms the state-of-the-art (ℓ1-ESPIRiT, calibration-less AC-LORAKS and CaLM methods) significantly on magnitude images for the T1 and FLAIR contrasts. Additionally, further validation operated on 8 to 20-fold prospectively accelerated high-resolution ex vivo human brain MRI data collected at 7 Tesla confirms the retrospective results. Overall, OSCAR-based regularization preserves phase information more accurately (both visually and quantitatively) compared to other approaches, an asset that can only be assessed on real prospective experiments.

International audience; Adaptive importance sampling (AIS) methods are increasingly used for the approximation of distributions and related intractable integrals in the context of Bayesian inference. Population Monte Carlo (PMC) algorithms are a subclass of AIS methods, widely used due to their ease in the adaptation. In this paper, we propose a novel algorithm that exploits the benefits of the PMC framework and includes more efficient adaptive mechanisms, exploiting geometric information of the target distribution. In particular, the novel algorithm adapts the location and scale parameters of a set of importance densities (proposals). At each iteration, the location parameters are adapted by combining a versatile resampling strategy (i.e., using the information of previous weighted samples) with an advanced optimization-based scheme. Local second-order information of the target distribution is incorporated through a preconditioning matrix acting as a scaling metric onto a gradient direction. A damped Newton approach is adopted to ensure robustness of the scheme. The resulting metric is also used to update the scale parameters of the proposals. We discuss several key theoretical foundations for the proposed approach. Finally, we show the successful performance of the proposed method in three numerical examples, involving challenging distributions.

Imaging flow cytometry has become a popular technology for bioparticle image analysis because of its capability of capturing thousands of images per second. Nevertheless, the vast number of images generated by imaging flow cytometry imposes great challenges for data analysis especially when the species have similar morphologies. In this work, we report a deep learning-enabled high-throughput system for predicting Cryptosporidium and Giardia in drinking water. This system combines imaging flow cytometry and an efficient artificial neural network called MCellNet, which achieves a classification accuracy >99.6%. The system can detect Cryptosporidium and Giardia with a sensitivity of 97.37% and a specificity of 99.95%. The high-speed analysis reaches 346 frames per second, outperforming the state-of-the-art deep learning algorithm MobileNetV2 in speed (251 frames per second) with a comparable classification accuracy. The reported system empowers rapid, accurate, and high throughput bioparticle detection in clinical diagnostics, environmental monitoring and other potential biosensing applications. Ministry of Education (MOE) National Research Foundation (NRF) Public Utilities Board (PUB) Submitted/Accepted version This work was supported by the Singapore National Research Foun-dation under the Competitive Research Program (NRF-CRP13-2014-01), the Incentive for Research & Innovation Scheme(PUB-1804-0082) administered by the PUB, and Ministry of Educa-tion Tier 1 RG39/19

International audience; The annual IEEE 5-Minute Video Clip Contest (5-MICC) was launched by the IEEE Signal Processing Society (SPS), and the selected topic for the competition at IEEE ICIP 2020 was “Fight the Pandemic.” The organizing committee selected three finalist videos and placed them online for public voting. The first one is about a visual analytic system for pandemic management, the second concerns machine learning screening of coronavirus disease (COVID-19) patients based on X-ray images, and the third deals with a COVID-19 test strip reader. Taking the public voting results from more than 800 participants into consideration, the panel of judges decided the final rankings of the three videos. In this article, we present an overview of the 5-MICC at ICIP 2020, describing the competition setup, the teams, and their approaches. We also share our experience and the feedback we received from the finalists.

This survey is on recent advancements in the intersection of physical modeling and machine learning. We focus on the modeling of nonlinear systems which are closer to electric motors. Survey on motor control and fault detection in operation of electric motors has been done., Comment: arXiv admin note: submission has been withdrawn by arXiv administrators due to inappropriate text overlap with external source

This work addresses the problem of analyzing multi-channel time series data %. In this paper, we by proposing an unsupervised fusion framework based on %the recently proposed convolutional transform learning. Each channel is processed by a separate 1D convolutional transform; the output of all the channels are fused by a fully connected layer of transform learning. The training procedure takes advantage of the proximal interpretation of activation functions. We apply the developed framework to multi-channel financial data for stock forecasting and trading. We compare our proposed formulation with benchmark deep time series analysis networks. The results show that our method yields considerably better results than those compared against., Comment: Accepted at EUSIPCO 2020

International audience; In this work, we propose a novel neural network architecture, called Adaptive Convolutional Neural Network (ACNN), which can be viewed as an intermediate solution between a standard convolutional network and a fully connected one. A constrained training strategy is developed to learn the parameters of such a network. The proposed algorithm allows us to control the Lipschitz constant of our ACNN to secure its robustness to adversarial noise. The resulting learning approach is evaluated for signal denoising based on a database of music recordings. Both qualitative and quantitative results show that the designed network is successful in removing Gaussian noise with unknown variance

Explainability of deep neural networks is one of the most challenging and interesting problems in the field. In this study, we investigate the topic focusing on the interpretability of deep learning-based registration methods. In particular, with the appropriate model architecture and using a simple linear projection, we decompose the encoding space, generating a new basis, and we empirically show that this basis captures various decomposed anatomically aware geometrical transformations. We perform experiments using two different datasets focusing on lungs and hippocampus MRI. We show that such an approach can decompose the highly convoluted latent spaces of registration pipelines in an orthogonal space with several interesting properties. We hope that this work could shed some light on a better understanding of deep learning-based registration methods., Comment: 13 pages, 5 figures + 3 figures in supplementary materials Accepted to DART 2021 workshop

Graph Neural Networks (GNNs) achieve significant performance for various learning tasks on geometric data due to the incorporation of graph structure into the learning of node representations, which renders their comprehension challenging. In this paper, we first propose a unified framework satisfied by most existing GNN explainers. Then, we introduce GraphSVX, a post hoc local model-agnostic explanation method specifically designed for GNNs. GraphSVX is a decomposition technique that captures the "fair" contribution of each feature and node towards the explained prediction by constructing a surrogate model on a perturbed dataset. It extends to graphs and ultimately provides as explanation the Shapley Values from game theory. Experiments on real-world and synthetic datasets demonstrate that GraphSVX achieves state-of-the-art performance compared to baseline models while presenting core theoretical and human-centric properties., Comment: ECML PKDD 2021

International audience; The problem of maximizing or minimizing the spreading in a social network has become more timely than ever with the advent of fake news and the coronavirus epidemic. The solution to this problem pertains to influence maximization algorithms that identify the right nodes to lockdown for epidemic containment, hire for viral marketing campaigns, block for online political propaganda etc. Though these algorithms have been developed for many years, the majority of the literature focuses on scalability issues and relaxing the method's assumptions. In the recent years, the emergence of new complementary data and more advanced machine learning methods for decision have guided part of the literature towards learning-based approaches. These can range from learning how information spreads over a network, to learning how to solve the combinatorial optimization problem itself. In this tutorial, we aim to dissentangle and clearly define the different tasks around learning for influence applications in social networks. More specifically, we start from traditional influence maximization algorithms, describe the need of influence estimation and delineate the state-of-the-art on influence and diffusion learning. Subsequently, we delve into the problem of learning while optimizing the influence spreading which is based on online learning algorithms. Finally, we describe the latest approaches on learning influence maximization with graph neural networks and deep reinforcement learning.

International audience; Reflectance confocal microscopy (RCM) is a powerful tool to visualize the skin layers at cellular resolution up to adepth of 200µm. A semi-quantitative score of skin aging from RCM images has been previously published, requiring visualassessment of the images by experienced dermatologists. In this article we propose new computer-basedmethods to automatically quantify the skin aging process on RCM image.

International audience; Multiple instance learning (MIL) is the preferred approach for whole slide image classification. However, most MIL approaches do not exploit the interdependencies of tiles extracted from a whole slide image, which could provide valuable cues for classification. This paper presents a novel MIL approach that exploits the spatial relationship of tiles for classifying whole slide images. To do so, a sparse map is built from tiles embeddings, and is then classified by a sparse-input CNN. It obtained state-of-the-art performance over popular MIL approaches on the classification of cancer subtype involving 10000 whole slide images. Our results suggest that the proposed approach might (i) improve the representation learning of instances and (ii) exploit the context of instance embeddings to enhance the classification performance. The code of this work is open-source at {github censored for review}.

International audience; This work addresses the problem of completing a partially observed matrix where the entries are either ones or zeroes. This is typically called one-bit matrix completion or binary matrix completion. In this problem, the association among the rows and among the columns can be modeled through graph Laplacians. Since the Laplacians cannot be computed from the incomplete matrix, they must be simultaneously estimated while completing the matrix. We model the problem as graph regularized binary matrix completion where the graphs need to be learnt from the data. We proposed an algorithm based on an alternating minimization scheme, taking advantage of an efficient proximity-based inner solver. The algorithm is applied to the problem of collaborative filtering. Experiments on benchmark datasets with state-of-the-art techniques in collaborative filtering show that the proposed method improves over the rest by a considerable margin.

International audience; Geometric priors have been shown to be useful in image segmentation to regularize results. For example, the classical Mumford-Shah functional uses region perimeter as prior. This has inspired much research in the last few decades, with classical approaches like the Rudin-Osher-Fatemi and most graph-cut formulations, which all use a weighted or binary perimeter prior. It has been observed that this prior is not suitable in many applications, for example for segmenting thin objects or some textures, which may have high perimeter/surface ratio. Mumford observed that an interesting prior for natural objects is the Euler Elastical model, which involves the squared curvature. In other areas of science, researchers have noticed that some physical binarization processes, like emulsion unmixing can be well-approximated by curvature-related flow like the Willmore flow. However, curvature-related flows are not easy to compute because curvature is difficult to estimate accurately, and the underlying optimisation processes are not convex. In this article, we propose to formulate a digital flow that approximates an Elastica-related flow using a multigrid-convergent curvature estimator, within a discrete variational framework. We also present an application of this model as a post-processing step to a segmentation framework.

We introduce a neural network architecture to solve inverse problems linked to a one-dimensional integral operator. This architecture is built by unfolding a forward-backward algorithm derived from the minimization of an objective function which consists of the sum of a data-fidelity function and a Tikhonov-type regularization function. The robustness of this inversion method with respect to a perturbation of the input is theoretically analyzed. Ensuring robustness is consistent with inverse problem theory since it guarantees both the continuity of the inversion method and its insensitivity to small noise. The latter is a critical property as deep neural networks have been shown to be vulnerable to adversarial perturbations. One of the main novelties of our work is to show that the proposed network is also robust to perturbations of its bias. In our architecture, the bias accounts for the observed data in the inverse problem. We apply our method to the inversion of Abel integral operators, which define a fractional integration involved in wide range of physical processes. The neural network is numerically implemented and tested to illustrate the efficiency of the method. Lipschitz constants after training are computed to measure the robustness of the neural networks., Comment: 35 pages, 10 figures

Highlights • An algorithm for automatic Covid-19 quantification based on 2D & 3D deep convolutional neural networks is presented. • A Covid-19-specific holistic, highly compact multi-omics signature integrating imaging/clinical/ biological data and associated comorbidities for automatic patient staging is presented and evaluated. • Short and Long-term prognosis for clinical resources optimization offering alternative/complementary means to facilitate triage for Covid-19 • Clinically-relevant quantification and staging tool validated by comparison with clinical experts is reported., Graphical abstract, Coronavirus disease 2019 (COVID-19) emerged in 2019 and disseminated around the world rapidly. Computed tomography (CT) imaging has been proven to be an important tool for screening, disease quantification and staging. The latter is of extreme importance for organizational anticipation (availability of intensive care unit beds, patient management planning) as well as to accelerate drug development through rapid, reproducible and quantified assessment of treatment response. Even if currently there are no specific guidelines for the staging of the patients, CT together with some clinical and biological biomarkers are used. In this study, we collected a multi-center cohort and we investigated the use of medical imaging and artificial intelligence for disease quantification, staging and outcome prediction. Our approach relies on automatic deep learning-based disease quantification using an ensemble of architectures, and a data-driven consensus for the staging and outcome prediction of the patients fusing imaging biomarkers with clinical and biological attributes. Highly promising results on multiple external/independent evaluation cohorts as well as comparisons with expert human readers demonstrate the potentials of our approach.

International audience; Acquiring ultrasound images of suspected lesion areas allows radiologists to monitor the cancer development of patients. The goal of this paper is to provide an automatic lesion segmentation tool for assisting them on the analysis of ultrasound images, by relying on recent neural network methods. Specifically, we perform a comparative study for the segmentation of 348 ultrasound image pairs acquired in 19 centers across France, displaying different tumor types. We show that, with a careful hyperparameter tuning, U-net outperforms other state-of-the-art networks, reaching a Dice coefficient of 0.929. We then propose to introduce group convolution into U-net architecture. This leads to a lightweight network named Lighter U-net @128 that achieves comparable segmentation performance with obviously reduced model size, hence paving the way for an embedded integration within hospital environment. We made our code publicly available for reproducibility purpose.

MotivationInvestigation of existing drugs is an effective alternative to discovery of new drugs for treating diseases. This task of drug re-positioning can be assisted by various kinds of computational methods to predict the best indication for a drug given the open-source biological datasets. Owing to the fact that similar drugs tend to have common pathways and disease indications, the association matrix is assumed to be of low-rank structure. Hence, the problem of drug-disease association prediction can been modelled as a low-rank matrix-completion problem.ResultsIn this work, we propose a novel matrix completion framework which makes use of the sideinformation associated with drugs/diseases for the prediction of drug-disease indications modelled as neighborhood graph: Graph regularized 1-bit matrix compeltion (GR1BMC). The algorithm is specially designed for binary data and uses parallel proximal algorithm to solve the aforesaid minimization problem taking into account all the constraints including the neighborhood graph incorporation and restricting predicted scores within the specified range. The results of the proposed algorithm have been validated on two standard drug-disease association databases (Fdataset and Cdataset) by evaluating the AUC across the 10-fold cross validation splits. The usage of the method is also evaluated through a case study where top 5 indications are predicted for novel drugs and diseases, which then are verified with the CTD database. The results of these experiments demonstrate the practical usage and superiority of the proposed approach over the benchmark methods.Contactaanchalm@iiitd.ac.in

International audience; Anti-angiogenic therapy with bevacizumab is a widely used therapeutic option for recurrent glioblastoma (GBM). Nevertheless, the therapeutic response remains highly heterogeneous among GBM patients with discordant outcomes. Recent data have shown that radiomics, an advanced recent imaging analysis method, can help to predict both prognosis and therapy in a multitude of solid tumours. The objective of this study was to identify novel biomarkers, extracted from MRI and clinical data, which could predict overall survival (OS) and progression-free survival (PFS) in GBM patients treated with bevacizumab using machine-learning algorithms. In a cohort of 194 recurrent GBM patients (age range 18–80), radiomics data from pre-treatment T2 FLAIR and gadolinium-injected MRI images along with clinical features were analysed. Binary classification models for OS at 9, 12, and 15 months were evaluated. Our classification models successfully stratified the OS. The AUCs were equal to 0.78, 0.85, and 0.76 on the test sets (0.79, 0.82, and 0.87 on the training sets) for the 9-, 12-, and 15-month endpoints, respectively. Regressions yielded a C-index of 0.64 (0.74) for OS and 0.57 (0.69) for PFS. These results suggest that radiomics could assist in the elaboration of a predictive model for treatment selection in recurrent GBM patients.

International audience; Many data science problems can be efficiently addressed by minimizing a cost function subject to various constraints. In this paper a new method for solving largescale constrained differentiable optimization problems is proposed. To account efficiently for a wide range of constraints, our approach embeds a subspace algorithm into an exterior penalty framework. The subspace strategy, combined with a Majoration-Minimization step search, takes great advantage of the smoothness of the penalized cost function. Assuming that the latter is convex, the convergence of our algorithm to a solution of the constrained optimization problem is proved. Numerical experiments carried out on a large-scale image restoration application show that the proposed method outperforms stateof-the-art algorithms in terms of computational time.

International audience; We consider the proximal gradient algorithm for solving penalized least-squares minimization problems arising in data science. This first-order algorithm is attractive due to its flexibility and minimal memory requirements allowing to tackle large-scale minimization problems involving non-smooth penalties. However, for problems such as X-ray computed tomography, the applicability of the algorithm is dominated by the cost of applying the forward linear operator and its adjoint at each iteration. In practice, the adjoint operator is thus often replaced by an alternative operator with the aim to reduce the overall computation burden and potentially improve conditioning issues. In this paper, we propose to analyze the effect of such an adjoint mismatch on the convergence of the proximal gradient algorithm in an infinite-dimensional setting, thus generalizing the existing results on PGA. We derive conditions on the stepsize and on the gradient of the smooth part of the objective function under which convergence of the algorithm to a fixed point is guaranteed. We also derive bounds on the error between this point and the solution to the original minimization problem. We illustrate our theoretical findings with two image reconstruction tasks in computed tomography.

Graph Representation Learning aims to embed nodes in a low-dimensional space. In this thesis, we tackle various challenging problems arising in the field. Firstly, we study how to leverage the inherent local community structure of graphs while learning node representations. We learn enhanced community-aware representations by combining the latent information with the embeddings. Moreover, we concentrate on the expressive- ness of node representations. We emphasize exponential family distributions to capture rich interaction patterns. We propose a model that combines random walks with kernelized matrix factorization. In the last part of the thesis, we study models balancing the trade-off between efficiency and accuracy. We propose a scalable embedding model which computes binary node representations., L'objectif principal de l'Apprentissage de Représentations sur Graphes est de plonger les nœuds dans un espace vectoriel de petite dimension. Dans cette thèse, nous abordons plusieurs enjeux dans le domaine. Tout d'abord, nous étudions comment exploiter l'existence de communautés structurelles locales inhérentes aux graphes tout en apprenant les représentations. Nous apprenons des représentations améliorées de la communauté en combinant les informations latentes avec les représentations. De plus, nous nous concentrons sur l'expressivité des représentations. Nous mettons l'accent sur les distributions de familles exponentielles pour saisir des modèles d'interaction riches. Nous proposons un modèle qui combine les marches aléatoires avec une matrice de factorisation sous forme de noyau. Dans la dernière partie de la thèse, nous étudions des modèles permettant un bon compromis entre efficacité et précision. Nous proposons un modèle évolutif qui calcule des représentations binaires.

Recent deep neural networks have shown superb performance in analyzing bioimages for disease diagnosis and bioparticle classification. Conventional deep neural networks use simple classifiers such as SoftMax to obtain highly accurate results. However, they have limitations in many practical applications that require both low false alarm rate and high recovery rate, e.g., rare bioparticle detection, in which the representative image data is hard to collect, the training data is imbalanced, and the input images in inference time could be different from the training images. Deep metric learning offers a better generatability by using distance information to model the similarity of the images and learning function maps from image pixels to a latent space, playing a vital role in rare object detection. In this paper, we propose a robust model based on a deep metric neural network for rare bioparticle (Cryptosporidium or Giardia) detection in drinking water. Experimental results showed that the deep metric neural network achieved a high accuracy of 99.86% in classification, 98.89% in precision rate, 99.16% in recall rate and zero false alarm rate. The reported model empowers imaging flow cytometry with capabilities of biomedical diagnosis, environmental monitoring, and other biosensing applications. Ministry of Education (MOE) National Research Foundation (NRF) Published version This work was supported by the Singapore National Research Foundation under the Competitive Research Program (NRFCRP13- 2014-01), Ministry of Education Tier 1 RG39/19, and the Singapore Ministry of Education (MOE) Tier 3 grant (MOE2017-T3-1-001).

International audience; The reconstruction of a volumetric image from Digital Breast Tomosynthesis (DBT) measurements is an ill-posed inverse problem, for which existing iterative regularized approaches can provide a good solution. However, the clinical task is somehow omitted in the derivation of those techniques, although it plays a primary role in the radiologist diagnosis. In this work, we address this issue by introducing a novel variational formulation for DBT reconstruction, tailored for a specific clinical task, namely the detection of microcalcifications. Our method aims at simultaneously enhancing the detectability performance and enabling a high-quality restoration of the background breast tissues. Our contribution is threefold. First, we introduce an original task-based reconstruction framework through the proposition of a detectability function inspired from mathematical model observers. Second, we propose a novel total-variation regularizer where the gradient field accounts for the different morphological contents of the imaged breast. Third, we integrate the two developed measures into a cost function, minimized thanks to a new form of the Majorize Minimize Memory Gradient (3MG) algorithm. We conduct a numerical comparison of the convergence speed of the proposed method with those of standard convex optimization algorithms. Experimental results show the interest of our DBT reconstruction approach, qualitatively and quantitatively.

We demonstrate the benefit of a novel laser strategy in multiphoton microscopy (MPM). The cheap, simple, and turn-key supercontinuum laser system with its spectral shaping module, constitutes an ideal approach for the one-shot microscopic imaging of many fluorophores without modification of the excitation parameters: central wavelength, spectral bandwidth, and average power. The polyvalence of the resulting multiplex-multiphoton microscopy (M-MPM) device is illustrated by images of many biomedical models from several origins (biological, medical, or vegetal), generated while keeping constant the spectral parameters of excitation. The resolution of the M-MPM device is quantified by a procedure of point-spread-function (PSF) assessment led by an original, robust, and reliable computational approach FIGARO. The estimated values for the PSF width for our M-MPM system are shown to be comparable to standard values found in optical microscopy. The simplification of the excitation system constitutes a significant instrumental progress in biomedical MPM, paving the way to the imaging of many fluorophores with a single shot of excitation without any modification of the lighting device. RESEARCH HIGHLIGHTS: A new solution of multiplex-multiphoton microscopy device is shown, resting on a supercontinuum laser. The one-shot excitation device has imaged biomedical and vegetal models. Our original computational strategy measures usual microscopy resolution.

The reconstruction of a volumetric image from Digital Breast Tomosynthesis (DBT)measurements is an ill-posed inverse problem, for which existing iterative regularizedapproaches can provide a good solution. However, the clinical task is somehow omittedin the derivation of those techniques, although it plays a primary role in the radiologistdiagnosis. In this work, we address this issue by introducing a novel variational formulationfor DBT reconstruction. Our approach is tailored for a specific clinical task, namely the detection of microcalcifications. Our method aims at simultaneously enhancing the detectionperformance and enabling a high-quality restoration of the background breast tissues.First, we propose an original approach aiming at enhancing the detectability of microcalcifications in DBT reconstruction. Thus, we formulate a detectability function inspired from mathematical model observers. Then, we integrate it in a cost function which is minimized for 3D reconstruction of DBT volumes. Experimental results demonstrate the interest of our approach in terms of microcalcification detectability.In a second part, we introduce the Spatially Adaptive Total Variation (SATV) as a new regularization strategy applied to DBT reconstruction, in addition to the detectability function. Hence, an original formulation for the weighted gradient field is introduced, that efficiently incorporates prior knowledge on the location of small objects. Then, we derive our SATV regularization, and incorporate it in our proposed 3D reconstruction approach for DBT. We carry out several experiments, in which SATV regularizer shows a promising improvement with respect to state-of-the-art regularization methods.Third, we investigate the application of Majorize Minimize Memory Gradient (3MG) algorithm to our proposed reconstruction approach. Thus, we suggest two numerical improvements to boost the speed of the reconstruction scheme. Then, we assess the numerical performance of 3MG by comparing the convergence speed of the proposed method with state-of-the-art convex optimization algorithms.The last part of this thesis is focused on the quantitative assessment of the contribution of our proposed DBT reconstruction. Thus, we conduct a visual experiment trial involving fourteen readers including nine radiologists with different levels of expertise and five GE Healthcare experts in mammography. According to specific visual criteria, the results show the outperformance of our proposed reconstruction approach over the standard non-regularized least squares solution., La reconstruction en tomosynthèse numérique du sein est considérée comme un problème inverse, pour lequel les méthodes itératives régularisées permettent de fournir une bonne qualité d'image. Bien que la tâche clinique joue un rôle crucial lors de l’examen des images par le radiologue, elle n'a pas été jusqu'à présent directement prise en compte dans le processus de reconstruction des images de tomosynthèse. Dans cette thèse, nous introduisons une nouvelle formulation variationnelle de la reconstruction des images en tomosynthèse numérique du sein qui intègre la tâche clinique du radiologue, notamment la détection des microcalcifications. Le but de cette approche est de permettre à la fois le rehaussement de la détectabilité des microcalcifications et une restauration de bonne qualité des tissus mammaires.Tout d'abord, nous proposons une nouvelle approche qui vise à rehausser la détectabilité des microcalcifications. Nous formulons une nouvelle fonction de détectabilité inspirée d’observateurs mathématiques. Nous l’intégrons, par la suite, dans une fonction objectif minimisée par un algorithme de reconstruction dédié. Nous montrons finalement l'intérêt de notre approche à l'égard des méthodes standards de reconstruction. Dans une deuxième partie, nous introduisons une nouvelle régularisation, Spatially Adaptive Total Variation (SATV), en complément de la fonction de détectabilité dans le problème de reconstruction en tomosynthèse. Nous proposons une formulation originale où l’opérateur de gradient est remplacé par un opérateur adaptatif appliqué à l'image qui incorpore efficacement la connaissance a priori relative à la localisation de petits objets. Ensuite, nous dérivons notre régularisation SATV et l'intégrons dans une nouvelle approche de reconstruction. Les résultats expérimentaux montrent que SATV est une piste prometteuse pour améliorer les méthodes de régularisation de l’état de l'art. Dans une troisième partie, nous étudions l'application de l’algorithme de Majoration-Minimisation à Mémoire de Gradient (3MG) à notre problème de reconstruction. Dans le but d’accroître sa vitesse de convergence, nous proposons deux améliorations numériques. Dès lors, les performances numériques sont évaluées en comparant la vitesse de convergence de la méthode proposée avec celles d'algorithmes d'optimisation convexe concurrents. La dernière partie de la thèse porte sur l'évaluation quantitative des contributions de l'approche de reconstruction proposée en tomosynthèse numérique du sein. Nous menons une étude de lecture d'images impliquant quatorze lecteurs dont neuf radiologues avec différents niveaux d'expertise et cinq experts en mammographie de GE Healthcare. Les résultats démontrent l'intérêt de notre approche de reconstruction par rapport à l'approche standard non-régularisée selon des critères visuels spécifiques.

A wide class of nonconvex optimization problem is represented by rational optimization problems. The latter appear naturally in many areas such as signal processing or chemical engineering. However, finding the global optima of such problems is intricate. A recent approach called Lasserre's hierarchy provides a sequence of convex problems that has the theoretical guarantee to converge to the global optima. Nevertheless, this approach is computationally challenging due to the high dimensions of the convex relaxations. In this thesis, we tackle this challenge for various signal processing problems.First, we formulate the reconstruction of sparse signals as a rational optimization problem. We show that the latter has a structure that we wan exploit in order to reduce the complexity of the associated relaxations. We thus solve several practical problems such as the reconstruction of chromatography signals. We also extend our method to the reconstruction of various types of signal corrupted by different noise models.In a second part, we study the convex relaxations generated by our problems which take the form of high-dimensional semi-definite programming problems. We consider several algorithms mainly based on proximal operators to solve those high-dimensional problems efficiently.The last part of this thesis is dedicated to the link between polynomial optimization and symmetric tensor decomposition. Indeed, they both can be seen as an instance of the moment problem. We thereby propose a detection method as well as a decomposition algorithm for symmetric tensors based on the tools used in polynomial optimization. In parallel, we suggest a robust extraction method for polynomial optimization based on tensor decomposition algorithms. Those methods are illustrated on signal processing problems.; Une vaste classe de problèmes d'optimisation non convexes est celle de l'optimisation rationnelle. Cette dernière apparaît naturellement dans de nombreux domaines tels que le traitement du signal ou le génie des procédés. Toutefois, trouver les optima globaux pour ces problèmes est difficile. Une approche récente, appelée la hiérarchie de Lasserre, fournit néanmoins une suite de problèmes convexes assurée de converger vers le minimum global. Cependant, cette approche représente un défi calculatoire du fait de la très grande dimension de ses relaxations. Dans cette thèse, nous abordons ce défi pour divers problèmes de traitement du signal.Dans un premier temps, nous formulons la reconstruction de signaux parcimonieux en un problème d'optimisation rationnelle. Nous montrons alors que ce dernier possède une structure que nous exploitons afin de réduire la complexité des relaxations associées. Nous pouvons ainsi résoudre plusieurs problèmes pratiques comme la restoration de signaux de chromatographie. Nous étendons également notre méthode à la restoration de signaux dans différents contextes en proposant plusieurs modèles de bruit et de signal. Dans une deuxième partie, nous étudions les relaxations convexes générées par nos problèmes et qui se présentent sous la forme de problèmes d'optimisation semi-définie positive de très grandes dimensions. Nous considérons plusieurs algorithmes basés sur les opérateurs proximaux pour les résoudre efficacement.La dernière partie de cette thèse est consacrée au lien entre les problèmes d'optimisation polynomiaux et la décomposition de tenseurs symétriques. En effet, ces derniers peuvent être tous deux vus comme une instance du problème des moments. Nous proposons ainsi une méthode de détection de rang et de décomposition pour les tenseurs symétriques basée sur les outils connus en optimisation polynomiale. Parallèlement, nous proposons une technique d'extraction robuste des solutions d'un problème d'optimisation poylnomiale basée sur les algorithmes de décomposition de tenseurs. Ces méthodes sont illustrées sur des problèmes de traitement du signal.

High accuracy measurement of size is essential in physical and biomedical sciences. Various sizing techniques have been widely used in sorting colloidal materials, analyzing bioparticles and monitoring the qualities of food and atmosphere. Most imaging-free methods such as light scattering measure the averaged size of particles and have difficulties in determining non-spherical particles. Imaging acquisition using camera is capable of observing individual nanoparticles in real time, but the accuracy is compromised by the image defocusing and instrumental calibration. In this work, a machine learning-based pipeline is developed to facilitate a high accuracy imaging-based particle sizing. The pipeline consists of an image segmentation module for cell identification and a machine learning model for accurate pixel-to-size conversion. The results manifest a significantly improved accuracy, showing great potential for a wide range of applications in environmental sensing, biomedical diagnostical, and material characterization. Ministry of Education (MOE) National Research Foundation (NRF) Published version This research was funded by the Singapore National Research Foundation under the Competitive Research Program (NRF-CRP13-2014-01), and Ministry of Education Tier 1 RG39/19.

International audience; This work introduces a new unsupervised representation learning technique called Deep Convolutional Transform Learning (DCTL). By stacking convolutional transforms, our approach is able to learn a set of independent kernels at different layers. The features extracted in an unsupervised manner can then be used to perform machine learning tasks, such as classification and clustering. The learning technique relies on a well-sounded alternating proximal minimization scheme with established convergence guarantees. Our experimental results show that the proposed DCTL technique outperforms its shallow version CTL, on several benchmark datasets.

This work proposes an unsupervised fusion framework based on deep convolutional transform learning. The great learning ability of convolutional filters for data analysis is well acknowledged. The success of convolutive features owes to convolutional neural network (CNN). However, CNN cannot perform learning tasks in an unsupervised fashion. In a recent work, we show that such shortcoming can be addressed by adopting a convolutional transform learning (CTL) approach, where convolutional filters are learnt in an unsupervised fashion. The present paper aims at (i) proposing a deep version of CTL; (ii) proposing an unsupervised fusion formulation taking advantage of the proposed deep CTL representation; (iii) developing a mathematically sounded optimization strategy for performing the learning task. We apply the proposed technique, named DeConFuse, on the problem of stock forecasting and trading. Comparison with state-of-the-art methods (based on CNN and long short-term memory network) shows the superiority of our method for performing a reliable feature extraction., Accepted at EURASIP JASP 2020

International audience; Registration of multi-modal medical images is an essential pre-processing step, for example, for fusion or image guided-interventions. However, the alignment process is prone to high variability in tissue appearance between modalities, in addition to local intensity variations and artefacts. This work introduces a robust multi-modal registration approach that mitigates the undesirable effect of such variability. Robustness is achieved using Huber's loss function for the data fidelity and regularization terms. We propose a novel approach using Huber's criterion, which enables a jointly convex estimation of the motions and the associated scale parameters. We formulate the problem as a complex 2D transformation estimation and investigate a robust total-variation smoothing, as well as a dictionary learning-based data fidelity term. Experiments are conducted using two datasets of multi-contrast MR brain images.

Component-graphs provide powerful and complex structures for multi-band image processing. We propose a multiband astronomical source detection framework with the component-graphs relying on a new set of component attributes. We propose two modules to differentiate nodes belong to distinct objects and to detect partial object nodes. Experiments demonstrate an improved capacity at detecting faint objects on a multi-band astronomical dataset.