Participants: Emilie Chouzenoux, Jean-Christophe Pesquet (Collaboration: Henri Gérard, ENPC, Paris)

A wide array of machine learning problems are formulated as the minimization of the expectation of a convex loss function on some parameter space. Since the probability distribution of the data of interest is usually unknown, it is is often estimated from training sets, which may lead to poor out-of-sample performance. In the work , we bring new insights in this problem by using the framework which has been developed in quantitative finance for risk measures. We show that the original min-max problem can be recast as a convex minimization problem under suitable assumptions. We discuss several important examples of robust formulations, in particular by defining ambiguity sets based on $\varphi$-divergences and the Wasserstein metric. We also propose an efficient algorithm for solving the corresponding convex optimization problems involving complex convex constraints. Through simulation examples, we demonstrate that this algorithm scales well on real data sets.

Participants: Emilie Chouzenoux (Collaboration: Aanchal Mongia, Neha Jhamb, Angshul Majumdar, IIIT Delhi, India)

Latent factor models have been used widely in collaborative filtering based recommender systems. In recent years, deep learning has been successful in solving a wide variety of machine learning problems. Motivated by the success of deep learning, we propose in , a deeper version of latent factor model. Experiments on benchmark datasets shows that our proposed technique significantly outperforms all stateof-the-art collaborative filtering techniques.

Participants: Emilie Chouzenoux, Marie-Caroline Corbineau, Jean-Christophe Pesquet

In the work , we introduce a new proximal interior point algorithm (PIPA). This algorithm is able to handle convex optimization problems involving various constraints where the objective function is the sum of a Lipschitz differentiable term and a possibly nonsmooth one. Each iteration of PIPA involves the minimization of a merit function evaluated for decaying values of a logarithmic barrier parameter. This inner minimization is performed thanks to a finite number of subiterations of a variable metric forward-backward method employing a line search strategy. The convergence of this latter step as well as the convergence the global method itself are analyzed. The numerical efficiency of the proposed approach is demonstrated in two image processing applications.

Participants: Emilie Chouzenoux, Marie-Caroline Corbineau, Jean-Christophe Pesquet (Collaboration: Carla Bertocchi, Marco Prato, Universita di Modena, Italy)

Variational methods are widely applied to ill-posed inverse problems for they have the ability to embed prior knowledge about the solution. However, the level of performance of these methods significantly depends on a set of parameters, which can be estimated through computationally expensive and time consuming methods. In contrast, deep learning offers very generic and efficient architectures, at the expense of explainability, since it is often used as a black-box, without any fine control over its output. Deep unfolding provides a convenient approach to combine variational-based and deep learning approaches. Starting from a variational formulation for image restoration, we developped in , , iRestNet, a neural network architecture obtained by unfolding a proximal interior point algorithm. Hard constraints, encoding desirable properties for the restored image, are incorporated into the network thanks to a logarithmic barrier, while the barrier parameter, the stepsize, and the penalization weight are learned by the network. We derive explicit expressions for the gradient of the proximity operator for various choices of constraints, which allows training iRestNet with gradient descent and backpropagation. In addition, we provide theoretical results regarding the stability of the network for a common inverse problem example. Numerical experiments on image deblurring problems show that the proposed approach compares favorably with both state-of-the-art variational and machine learning methods in terms of image quality.

Participants: Emilie Chouzenoux, Marie-Caroline Corbineau, Jean-Christophe Pesquet (Collaboration: Denis Kouamé, Jean-Yves Tourneret, IRIT, Toulouse)

Joint deconvolution and segmentation of ultrasound images is a challenging problem in medical imaging. By adopting a hierarchical Bayesian model, we propose in an accelerated Markov chain Monte Carlo scheme where the tissue reflectivity function is sampled thanks to a recently introduced proximal unadjusted Langevin algorithm. This new approach is combined with a forward-backward step and a preconditioning strategy to accelerate the convergence, and with a method based on the majorization-minimization principle to solve the inner nonconvex minimization problems. As demonstrated in numerical experiments conducted on both simulated and in vivo ultrasound images, the proposed method provides high-quality restoration and segmentation results and is up to six times faster than an existing Hamiltonian Monte Carlo method.

Participants: Emilie Chouzenoux, Jean-Christophe Pesquet (Collaboration: Giovanni Chierchia, ESIEE Paris, Luis Bricenos Arias, Universidad Técnica Federico Santa Maria, Valparaiso, Chile)

In the paper , we proposed a new optimization algorithm for sparse logistic regression based on a stochastic version of the Douglas-Rachford splitting method. Our algorithm sweeps the training set by randomly selecting a mini-batch of data at each iteration, and it allows us to update the variables in a block coordinate manner. Our approach leverages the proximity operator of the logistic loss, which is expressed with the generalized Lambert W function. Experiments carried out on standard datasets demonstrate the efficiency of our approach w.r.t. stochastic gradient-like methods.

Participant: Emilie Chouzenoux (Collaboration: Omer Deniz Akyildiz, Alan Turing Institute, London, UK, Victor Elvira, University of Edinburgh, Joaquin Miguez, Universidad Carlos III de Madrid, Spain)

In the paper , we proposed a probabilistic optimization method, named probabilistic incremental proximal gradient (PIPG) method, by developing a probabilistic interpretation of the incremental proximal gradient algorithm. We explicitly model the update rules of the incremental proximal gradient method and develop a systematic approach to propagate the uncertainty of the solution estimate over iterations. The PIPG algorithm takes the form of Bayesian filtering updates for a state-space model constructed by using the cost function. Our framework makes it possible to utilize well-known exact or approximate Bayesian filters, such as Kalman or extended Kalman filters, to solve large scale regularized optimization problems.

Participants: Emilie Chouzenoux, Jean-Christophe Pesquet (Collaboration: Claire Lefort, XLIM, Limoges, Tim Tsz-Kit Lau, Northwestern University, USA)

Fitting Gaussian functions to empirical data is a crucial task in a variety of scientific applications, especially in image processing. However, most of the existing approaches for performing such fitting are restricted to two dimensions and they cannot be easily extended to higher dimensions. Moreover, they are usually based on alternating minimization schemes which benefit from few theoretical guarantees in the underlying nonconvex setting. In the paper , we provided a novel variational formulation of the multivariate Gaussian fitting problem, which is applicable to any dimension and accounts for possible non-zero background and noise in the input data. The block multiconvexity of our objective function leads us to propose a proximal alternating method to minimize it in order to estimate the Gaussian shape parameters. The resulting FIGARO algorithm is shown to converge to a critical point under mild assumptions. The algorithm shows a good robustness when tested on synthetic datasets. To demonstrate the versatility of FIGARO, we also illustrate its excellent performance in the fitting of the Point Spread Functions of experimental raw data from a two-photon fluorescence microscope.

Participants: Emilie Chouzenoux, Loubna El Gueddari (Collaboration: Philippe Ciuciu, Alexandre Vignaut, Inria Saclay, Parietal)

Over the last decade, the combination of parallel imaging (PI) and compressed sensing (CS) in magnetic resonance imaging (MRI) has allowed to speed up acquisition while maintaining a good signal-to-noise ratio (SNR) for millimetric resolution. Self-calibrating techniques such as L1-ESPiRIT have emerged as a standard approach to estimate the coil sensitivity maps that are required at the reconstruction stage. Although straightforward in Cartesian acquisitions, these approaches become more computationally demanding in non-Cartesian scenarios especially for high resolution imaging (e.g. 500 $\mu$m in plane). Instead, calibration-less techniques no longer require this prior knowledge to perform multi-channel image reconstruction from undersampled k-space data. In this work, we introduce a new calibration-less PI-CS reconstruction method that is particularly suited to non-Cartesian data. It leverages structure sparsity of the multi-channel images in a wavelet transform domain while adapting to SNR inhomogeneities across receivers thanks to the OSCAR-norm regularization. Comparison and validation on 8 to 20-fold prospectively accelerated high-resolution ex-vivo human brain MRI data collected at 7 Tesla shows that the subbandwise OSCAR-norm regularization achieves the best trade-off between image quality and computational cost at the reconstructions stage compared to other tested versions (global, scalewise and pixelwise). This approach provides slight to moderate improvement over its state-of-the-art competitors (self-calibrating `1-ESPIRiT method and calibration-less AC-LORAKS and CaLM methods) in terms of closeness to the Cartesian reference magnitude image. Importantly, it also preserves much better phase information compared to other approaches , , .

Participants: Emilie Chouzenoux, Jean-Christophe Pesquet (Collaboration: Alessandro Benfenati, Universita di Milano)

In recent years, there has been a growing interest in mathematical models leading to the minimization, in a symmetric matrix space, of a Bregman divergence coupled with a regularization term. We address problems of this type within a general framework where the regularization term is split into two parts, one being a spectral function while the other is arbitrary. A Douglas–Rachford approach is proposed to address such problems, and a list of proximity operators is provided allowing us to consider various choices for the fit–to–data functional and for the regularization term. Based on our theoretical results, two novel approaches are proposed for the noisy graphical lasso problem, where a covariance or precision matrix has to be statistically estimated in presence of noise. The Douglas–Rachford approach directly applies to the estimation of the covariance matrix. When the precision matrix is sought, we solve a nonconvex optimization problem. More precisely, we propose a majorization–minimization approach building a sequence of convex surrogates and solving the inner optimization subproblems via the aforementioned Douglas–Rachford procedure. We establish conditions for the convergence of this iterative scheme. We illustrate the good numerical performance of the proposed approaches with respect to state–of–the–art approaches on synthetic and real-world datasets .

Participants: Fragkiskos Malliaros, Abdulkadir Çelikkanat

Network representation learning (NRL) methods aim to map each vertex into a low dimensional space by preserving both local and global structure of a given network. In recent years, various approaches based on random walks have been proposed to learn node embeddings – thanks to their success in several challenging problems. In this work, we have introduced two methodologies to compute latent representations of nodes based on random walks.

In particular, we have proposed Kernel Node Embeddings (KernelNE) , a model that aims to bring together two popular approaches for NRL, namely matrix factorization and random walk-based models. KernelNE is a weighted matrix factorization model which encodes random walk-based information about the nodes of the graph. The main benefit of this formulation is that it allows to utilize kernel functions on the computation of the embeddings.

Our second approach is motivated by the fact that the popular Skip-Gram algorithm models the conditional distribution of nodes within a random walk based on the softmax function, which might prohibit to capture richer types of interaction patterns among nodes that co-occur within a random walk. Here we argue that considering more expressive conditional probability models to relate nodes within a random walk sequence, might lead to more informative representations. That way, we have introduced the Exponential Family Graph Embedding (EFGE) model , that capitalizes on exponential family distribution models to capture interactions between nodes.

We have evaluated our methods on two downstream tasks: node classification and link prediction in social, information and biological networks. The experimental results demonstrate that random walk-based models accompanied with kernels as well as exponential family distributions outperform widely-known baseline NRL methods.

Participants: Fragkiskos Malliaros (Collaboration: Adrien Benamira, Benjamin Devillers, Etienne Lesot, Ayush K. Ray, Manal Saadi, CentraleSupélec)

Social networks have become the main platforms for information dissemination. Nevertheless, due to the increasing number of users, social media platforms tend to be highly vulnerable to the propagation of disinformation – making the detection of fake news a challenging task. In our work, we have focused on content-based methods for detecting fake news – casting the problem to a binary text classification one (an article corresponds to either fake news or not). The main challenge here stems from the fact that the number of labeled data is limited; very few articles can be examined and annotated as fake. To this extend, we opted for semi-supervised learning approaches. In particular, we have proposed a graph-based semi-supervised fake news detection method, based on graph neural networks . Our intuition is that, graphs are expressive models that are able to capture contextual dependencies among articles, alleviating the label scarcity constraint. On a high level, our framework is composed of three components: $i$) embedding of articles in the Euclidean space; $ii$) construction of an article similarity graph; $iii$) inference of missing labels using graph learning techniques. The experimental results indicate that the proposed methodology achieves better performance compared to traditional classification techniques, especially when trained on limited number of labeled articles.

Participants: Fragkiskos Malliaros (Collaboration: Antoine J.-P. Tixier, Maria Evgenia G. Rossi, Jesse Read, Michalis Vazirgiannis, École Polytechnique)

Influential spreaders are nodes that can diffuse information to the largest part of the network in a minimum amount of time. Detecting influential spreaders is an important task with numerous real-world applications. Nevertheless, some of the most effective influential spreader detection algorithms (e.g., the $k$-core decomposition) are unstable to small perturbations of the network structure. Inspired by bagging in Machine Learning, we have proposed the first Perturb and Combine (P&C) procedure for networks . It (1) creates many perturbed versions of a given graph, (2) applies a node scoring function separately to each graph, and (3) combines the results. Experiments conducted on real-world networks of various sizes with the $k$-core, generalized $k$-core, and PageRank algorithms reveal that P&C brings substantial improvements. Moreover, this performance boost can be obtained at almost no extra cost through parallelization. Finally, a bias-variance analysis suggests that P&C works mainly by reducing bias, and that therefore, it should be capable of improving the performance of all vertex scoring functions, including stable ones.

Participant: Jean-Christophe Pesquet (Collaboration: Patrick Louis Combettes, North Carolina University, USA)

In our previous work, we investigated the almost sure weak convergence of block-coordinate fixed point algorithms and discussed their applications to nonlinear analysis and optimization. This algorithmic framework features random sweeping rules to select arbitrarily the blocks of variables that are activated over the course of the iterations and it allows for stochastic errors in the evaluation of the operators. The paper establishes results on the mean-square and linear convergence of the iterates. Applications to monotone operator splitting and proximal optimization algorithms are presented.

Participants: Marc Castella, Arthur Marmin, Jean-Christophe Pesquet

Recovering nonlinearly degraded signal in the presence of noise is a challenging problem. In this work, this problem is tackled by minimizing the sum of a non convex least-squares fit criterion and a penalty term. We assume that the nonlinearity of the model can be accounted for by a rational function. In addition, we suppose that the signal to be sought is sparse and a rational approximation of the ${\ell }_0$ pseudo-norm thus constitutes a suitable penalization. The resulting composite cost function belongs to the broad class of semi-algebraic functions. To find a globally optimal solution to such an optimization problem, it can be transformed into a generalized moment problem, for which a hierarchy of semidefinite programming relaxations can be built. Global optimality comes at the expense of an increased dimension and, to overcome computational limitations concerning the number of involved variables, the structure of the problem has to be carefully addressed. A situation of practical interest is when the nonlinear model consists of a convolutive transform followed by a componentwise nonlinear rational saturation. We then propose to use a sparse relaxation able to deal with up to several hundreds of optimized variables. In contrast with the naive approach consisting of linearizing the model, our experiments show that the proposed approach offers good performance .

Participant: Jean-Christophe Pesquet (Collaboration: Patrick Louis Combettes, North Carolina University, USA)

Motivated by structures that appear in deep neural networks, we investigate nonlinear composite models alternating proximity and affine operators defined on different spaces. We first show that a wide range of activation operators used in neural networks are actually proximity operators. We then establish conditions for the averagedness of the proposed composite constructs and investigate their asymptotic properties. It is shown that the limit of the resulting process solves a variational inequality which, in general, does not derive from a minimization problem .

Participant: Hugues Talbot (Collaboration: Clara Jaquet, Laurent Najman, ESIEE Paris, Leo Grady, Michiel Schaap, Buzzy Spain, Hyun Kim, Charles Taylor, HeartFlow Inc, Irene Vignon-Clementel, Inria Paris)

In this work, we have proposed a blood-vessel generation procedure for extending known patient vasculature over and within the heart ventricle . It is patient-specific, in the sense that it extend the known, segmented patient vasculature, and it is consistent with physics-based blood vessels characteristics (i.e. derived from CFD) and known vessel physiology. The generated vascular network bridges the gap between the vasculature that can be imaged and assessed via classical means (CT or MRI) and perfusion maps that can be imaged with specific modalities (radiotracer injected scintigraphy or PET). One objective of this work is to eventually propose a forward model for perfusion map generation, that can be used to solved the associated inverse problem of finding the cause of observed perfusion deficits associated with coronary diseases that cannot be imaged directly.

Participant: Hugues Talbot (Collaboration: Diane Genest, Élodie Puybareau, Jean Cousty, ESIEE Paris, Marc Léonard, Noémie de Crozé, L'Oréal Recherche)

Fish embryos are used throughout the cosmetics industry to assess the toxicity of the components of their products, as well as more generally in waterways pollution measurements. Indeed pollution is often detectable in trace amounts when they hinder, stop or cause malformations during fish embryo development. In this work, we propose a high-throughput procedure for detecting most important malformations in fish embryo. For examples those affecting the tail or the eyes, based on image analysis and machine learning . We have also proposed an atlas-based automated procedure for detecting swim bladder malformations, which are very difficult to assess manually .

These malformation are among the most difficult to assess but very common in various degrees of severity. Our procedure provide similar error rate as trained and careful humans operators, as assessed on thousands of images acquired in partneship with L'Oréal. We also show that our procedure is much faster and more consistent than human operators. It is now used in production by our partner.

Participant: Hugues Talbot (Collaboration: Eloïse Grossiord, Laurent Najman, ESIEE Paris, Benoît Naegel, iCube, Strasbourg, Nicolas Passat, CRESTIC, Reims)

Lymphoma is type of cancer affecting the lymph system. Similar to blood disorders, these cancers can be difficult to cure because they affect a large portion of the body and metastasize easily. In contrast to leukemia, lymphoma also affects organs: the lymph nodes. Assessing the effectiveness of therapies implies to follow the impact of treatment on lymph nodes. This requires segmenting a large number of lesion, often several dozens. In , we have proposed an automated procedure based on hierarchical mathematical morphology, which has been extensively validated, and is now available as a plug-in for ImageJ/FIJI.

Participant: Hugues Talbot (Collaboration: Odyssée Merveille, Benoît Naegel, iCube, Strasbourg, Nicolas Passat, CRESTIC, Reims)

Curvilinear structure restoration in image processing procedures is a difficult task, which can be compounded when these structures are thin, i.e., when their smallest dimension is close to the resolution of the sensor. Many recent restoration methods involve considering a local gradient-based regularization term as prior, assuming gradient sparsity. An isotropic gradient operator is typically not suitable for thin curvilinear structures, since gradients are not sparse for these. In this paper , we propose a mixed gradient operator that combines a standard gradient in the isotropic image regions, and a directional gradient in the regions where specific orientations are likely. In particular, such information can be provided by curvilinear structure detectors (e.g., RORPO or Frangi filters). Our proposed mixed gradient operator, that can be viewed as a companion tool of such detectors, is proposed in a discrete framework and its formulation/computation holds in any dimension; in other words, it is valid in $Z^n$, $n\ge 1$. We show how this mixed gradient can be used to construct image priors that take edge orientation, as well as intensity, into account, and then involved in various image processing tasks while preserving curvilinear structures. The experiments carried out on 2D, 3D, real, and synthetic images illustrate the relevance of the proposed gradient, and its use in variational frameworks for both denoising and segmentation tasks.

Participant: Hugues Talbot (Collaboration: Julie Robic, Alex Nkengne, Clarins laboratories, Benjamin Perret, Michel Couprie, ESIEE Paris)

With aging, human skin becomes drier, thinner and more irregular, but these characteristics are highly person-dependent, and can be brought about via exposure to heat, cold or Sun. It is important to the cosmetics industry to assess objectively the effect of their products on skin aging. With our partner Clarins laboratory, we have proposed a series of automated procedures based on graph-based image analysis. We have in particular proposed to detect the surface that correspond to the dermal-epidermal junction , and a series or procedures to link the appearance of this surfact to aging characteristics . Both have been validated by dermatologists specialized in skin aging.

Participant: Hugues Talbot (Collaboration: Alessandro Benfenati, Universita di Milano, Francesco Bonacci, Laboratoire Navier, Tarik Bourouina, ESIEE Paris)

Fluorescent bead tracking is important in biomedical application related to biomechanics, rheology and fluid dynamics. We have made several contributions for the detection and tracking of micrometer-scale fluorescent bead in 3D confocal microscopy , . Many software packages exist for 2D tracking but almost none exist for 3D. It is a harder problem because in general beads are not fixed and move between plane acquisitions, due to the relatively slow scanning characteristics of confocal microscopy.

Participants: Guillaume Chassagnon, Maria Vakalopoulou (Collaboration: Marie-Pierre Revel and Nikos Paragios, AP-HP - Hopital Cochin Broca Hotel Dieu, Therapanacea)

Relevance and penetration of machine learning in clinical practice is a recent phenomenon with multiple applications being currently under development. Deep learning –and especially convolutional neural networks (CNNs)– is a subset of machine learning, which has recently entered the field of thoracic imaging. The structure of neural networks, organized in multiple layers, allows them to address complex tasks. For several clinical situations, CNNs have demonstrated superior performance as compared with classical machine learning algorithms and in some cases achieved comparable or better performance than clinical experts. Chest radiography, a high-volume procedure, is a natural application domain because of the large amount of stored images and reports facilitating the training of deep learning algorithms. Several algorithms for automated reporting have been developed. The training of deep learning algorithm CT images is more complex due to the dimension, variability, and complexity of the 3D signal. The role of these methods is likely to increase in clinical practice as a complement of the radiologist's expertise. The objective of these two reviews , is to provide definitions for understanding the methods and their potential applications for thoracic imaging.

Participants: Guillaume Chassagnon, Maria Vakalopoulou (Collaboration: Charlotte Martin, Rafael Marini Silva, Alexis Régent, Luc Mouthon, Nikos Paragios and Marie-Pierre Revel, AP-HP - Hopital Cochin Broca Hotel Dieu, Therapanacea)

Elastic registration of inspiratory and expiratory MRI revealed qualitative and quantitative differences in lung deformation in study participants with systemic sclerosis compared with healthy volunteers. Current imaging methods are not sensitive to changes in pulmonary function resulting from fibrosis. MRI with ultrashort echo time can be used to image the lung parenchyma and lung motion. To evaluate elastic registration of inspiratory to expiratory lung MRI for the assessment of pulmonary fibrosis in study participants with systemic sclerosis (SSc). This prospective study was performed from September 2017 to March 2018 and recruited healthy volunteers and participants with SSc and high-resolution CT (within the previous 3 months) of the chest for lung MRI. Two breath-hold, coronal, three-dimensional, ultrashort–echo-time, gradient-echo sequences of the lungs were acquired after full inspiration and expiration with a 3.0-T unit. Images were registered from inspiration to expiration by using an elastic registration algorithm. Jacobian determinants were calculated from deformation fields and represented on color maps. Similarity between areas with marked shrinkage and logarithm of Jacobian determinants were compared between healthy volunteers and study participants with SSc. Receiver operating characteristic curve analysis was performed to determine the best Dice similarity coefficient threshold for diagnosis of fibrosis. Sixteen participants with SSc (seven with pulmonary fibrosis at high-resolution CT) and 11 healthy volunteers were evaluated. Areas of marked shrinkage during expiration with logarithm of Jacobian determinants less than -0.15 were found in the posterior lung bases of healthy volunteers and in participants with SSc without fibrosis, but not in participants with fibrosis. The sensitivity and specificity of MRI for presence of fibrosis at high-resolution CT were 86% and 75%, respectively (area under the curve, 0.81; P = .04) by using a threshold of 0.36 for Dice similarity coefficient. Elastic registration of inspiratory to expiratory MRI shows less lung base respiratory deformation in study participants with systemic sclerosis related pulmonary fibrosis compared with participants without fibrosis.

Participants: Théo Estienne, Enzo Battistella, Marvin Lerousseau, Roger Sun, Maria Vakalopoulou (Collaboration: Stergios Christodoulidis, Alexandre Carre, Guillaume Klausner, Stavroula Mougiakakou, Charlotte Robert, Nikos Paragios and Eric Deutsch, Institute Gustave Roussy, University of Bern, Therapanacea)

We proposed in a 3D deep neural network called U-ReSNet, a joint framework that can accurately register and segment medical volumes. The proposed network learns to automatically generate linear and elastic deformation models, trained by minimizing the mean square error and the local cross correlation similarity metrics. In parallel, a coupled architecture is integrated, seeking to provide segmentation maps for anatomies or tissue patterns using an additional decoder part trained with the dice coefficient metric. U-ReSNet is trained in an end to end fashion, while due to this joint optimization the generated network features are more informative leading to promising results compared to other deep learning-based methods existing in the literature. We evaluated the proposed architecture using the publicly available OASIS 3 dataset, measuring the dice coefficient metric for both registration and segmentation tasks. Our promising results indicate the potentials of our method which is composed from a convolutional architecture that is extremely simple and light in terms of parameters.

Participants: Enzo Battistella, Théo Estienne, Marvin Lerousseau, Roger Sun, Maria Vakalopoulou (Collaboration: Charlotte Robert, Nikos Paragios and Eric Deutsch, Institute Gustave Roussy, Therapanacea)

Precision medicine, a highly disruptive paradigm shift in healthcare targeting the personalizing treatment, heavily relies on genomic data. However, the complexity of the biological interactions, the important number of genes as well as the lack of substantial patient's clinical data consist a tremendous bottleneck on the clinical implementation of precision medicine. In this work , we introduce a generic, low dimensional gene signature that represents adequately the tumor type. Our gene signature is produced using LP-stability algorithm, a high dimensional center-based unsupervised clustering algorithm working in the dual domain, and is very versatile as it can consider any arbitrary distance metric between genes. The gene signature produced by LP-stability reports at least 10 times better statistical significance and 35% better biological significance than the ones produced by two referential unsupervised clustering methods. Moreover, our experiments demonstrate that our low dimensional biomarker (27 genes) surpass significantly existing state of the art methods both in terms of qualitative and quantitative assessment while providing better associations to tumor types than methods widely used in the literature that rely on several omics data.

Participants: Maria Papadomanolaki and Maria Vakalopoulou (Collaboration: Konstantinos Karantzalos, National Technical University of Athens)

Deep learning architectures have received much attention in recent years demonstrating state-of-the-art performance in several segmentation, classification and other computer vision tasks. Most of these deep networks are based on either convolutional or fully convolutional architectures. In this study , we propose a novel object-based deep-learning framework for semantic segmentation in very high-resolution satellite data. In particular, we exploit object-based priors integrated into a fully convolutional neural network by incorporating an anisotropic diffusion data preprocessing step and an additional loss term during the training process. Under this constrained framework, the goal is to enforce pixels that belong to the same object to be classified at the same semantic category. We compared thoroughly the novel object-based framework with the currently dominating convolutional and fully convolutional deep networks. In particular, numerous experiments were conducted on the publicly available ISPRS WGII/4 benchmark datasets, namely Vaihingen and Potsdam, for validation and inter-comparison based on a variety of metrics. Quantitatively, experimental results indicate that, overall, the proposed object-based framework slightly outperformed the current state-of-the-art fully convolutional networks by more than 1% in terms of overall accuracy, while intersection over union results are improved for all semantic categories. Qualitatively, man-made classes with more strict geometry such as buildings were the ones that benefit most from our method, especially along object boundaries, highlighting the great potential of the developed approach.

Participants: Maria Papadomanolaki and Maria Vakalopoulou (Collaboration: Konstantinos Karantzalos, National Technical University of Athens)

Semantic segmentation, especially for very high-resolution satellite data, is one of the pillar problems in the remote sensing community. Lately, deep learning techniques are the ones that set the state-of-the-art for a number of benchmark datasets, however, there are still a lot of challenges that need to be addressed, especially in the case of limited annotations. To this end, in this study , we propose a novel framework based on deep neural networks that is able to address concurrently semantic segmentation and image reconstruction in an end to end training. Under the proposed formulation, the image reconstruction acts as a regularization, constraining efficiently the solution in the entire image domain. This self-supervised component helps significantly the generalization of the network for the semantic segmentation, especially in cases of a low number of annotations. Experimental results and the performed quantitative evaluation on the publicly available ISPRS (WGIII/4) dataset indicate the great potential of the developed approach.

Participants: Maria Papadomanolaki, Sagar Verma and Maria Vakalopoulou (Collaboration: Siddharth Gupta and Konstantinos Karantzalos, Granular AI and National Technical University of Athens)

The advent of multitemporal high resolution data, like the Copernicus Sentinel-2, has enhanced significantly the potential of monitoring the earth's surface and environmental dynamics. In this study , we present a novel deep learning framework for urban change detection which combines state-of-the-art fully convolutional networks (similar to U-Net) for feature representation and powerful recurrent networks (such as LSTMs) for temporal modeling. We report our results on the recently publicly available bi-temporal Onera Satellite Change Detection (OSCD) Sentinel-2 dataset, enhancing the temporal information with additional images of the same region on different dates. Moreover, we evaluate the performance of the recurrent networks as well as the use of the additional dates on the unseen test-set using an ensemble cross-validation strategy. All the developed models during the validation phase have scored an overall accuracy of more than 95%, while the use of LSTMs and further temporal information, boost the F1 rate of the change class by an additional 1.5%.

Participants: Maria Vakalopoulou and Mihir Sahasrabudhe (Collaboration: Stergios Christodoulidis, Stavroula Mougiakakou and Nikos Paragios, University of Bern and Therapanacea)

Image registration in multimodal, multitemporal satellite imagery is one of the most important problems in remote sensing and essential for a number of other tasks such as change detection and image fusion. In this study , inspired by the recent success of deep learning approaches we propose a novel convolutional neural network architecture that couples linear and deformable approaches for accurate alignment of remote sensing imagery. The proposed method is completely unsupervised, ensures smooth displacement fields and provides real time registration on a pair of images. We evaluate the performance of our method using a challenging multitemporal dataset of very high resolution satellite images and compare its performance with a state of the art elastic registration method based on graphical models. Both quantitative and qualitative results prove the high potentials of our method.

Participants: Mihir Sahasrabudhe (Collaboration: Eqward Bartrum, Riza Alp Guler, Dimitris Samaras and Iasonas Kokkinos, Stony Brook, UCL and Ariel AI)

We introduced, in , Lifting Autoencoders, a generative 3D surface-based model of object categories. We bring together ideas from non-rigid structure from motion, image formation, and morphable models to learn a controllable, geometric model of 3D categories in an entirely unsupervised manner from an unstructured set of images. We exploit the 3D geometric nature of our model and use normal information to disentangle appearance into illumination, shading, and albedo. We further use weak supervision to disentangle the non-rigid shape variability of human faces into identity and expression. We combine the 3D representation with a differentiable renderer to generate RGB images and append an adversarially trained refinement network to obtain sharp, photorealistic image reconstruction results. The learned generative model can be controlled in terms of interpretable geometry and appearance factors, allowing us to perform photorealistic image manipulation of identity, expression, 3D pose, and illumination properties.