 by your business career, mostly in hand, in a particular place, is where he developed not only interesting research, but also contributed to tools that are used for the, like, Mikaï. Okay, thank you, Mikaï, for this nice introduction. So you asked me to make a presentation on addressing different topics, so I hope you won't be too much disparate for you, but so I will talk about the core of the presentation. It's on imaging biomarkers, and I will introduce some notions about algorithms for imaging biomarkers in the brain, and issue also on open data and open data for digital infrastructures using images. And because I am also a founding editor of a new series of frontiers in ICT called Image Computer Analysis, I will first maybe give you a rapid snapshot of what is this journal and why we are funding this. I was asked, in fact, to, okay, which is also a journal promoting open access for the data and for the publication. So I don't know if you're aware of the frontiers. Frontiers is mostly for frontiers and other things, so they launch a new section which is called Frontiers in ICT now, and it's funded by EPFL, which is around more than, almost 10 years ago. And two years ago, they have launched Frontiers in ICT, where they asked people to promote some different topics. So what is a particularity of frontiers? It's a real-time peer-review process that means it's interactive. It can be, it should be transparent. That means when your paper is published in frontiers, all remarks from the reviewers and even reviewers' names are disclosed unless they don't want to be disclosed. But otherwise, if they accept to be disclosed when a paper is published, every discussion between the authors and the reviewers is disclosed, which is also very interesting for all of the discussion and debates about the quality of the work. So the course at work, last editor, that's you to choose if it's true or not. But it's a quality-open access because it's sponsored by Nature Publishing Group, so they have a special agreement together. It's, it's time to have a very fast publication, about three months after submissions. Of course, it's according to the type of peer-review process. And we have, we have found that there are different, difference of culture between what people do mostly in biological science or neuroscience and what we used to do in ICT where we have more traditions of having a deep review of papers, which is not always the case in, in other, in other field. We have also post-publication evaluation in order to be able to have analytics of how papers are viewed, how papers are cited. And so the goal is to have the best paper to be published in this domain and it's indexed in most of the major database for publication. So Frontiers in ICT, there are different topics which have been launched. I don't know if there are new, more recent, so I'm leader with Patrick Boutemi of Computer Image Analytics because I think I'm not covering everything in Computer Image Analytics and I don't know if people of you know Patrick Boutemi but he's very well known in the field of image processing for motion detection on images and we are working at the same place so we are happy to do that together. Otherwise, you have other type of topics and when you submit a paper in Frontiers in ICT, you can ask your paper to be attached to one or several sub-topics according to what you do in the paper. And you can also, I will talk about that afterwards, have some special issue which can be also attached to different domains like we had, the special issue attached both in ICT and in Frontiers in North Science. So the main topics of Computer Image Analytics is to promote papers on image processing, image analysis and computer visions, emerge perceptions and all related application fields with a specific attention on emerging interdisciplinary. So we have a bunch of basic research domain which covers mostly probably the type of conference when people working in images are attending and also a larger set of applications and of course it's far to be only on life science, it could be environmental science, multimedia and video surveillance of biometrics. We have Editorial Board which is extended over time because as soon as you are reviewing in this journal, you become member of the Editorial Board which is also important for the contribution and there is also one important aspect which is called Research Topic and we are really dedicated to extend the notion of research topic which you can have a special issue but with some difference with special issues. The idea is to have a research topic with dedicated associated editor which manage the research topic that is peer reviewed and publish as soon as each paper is accepted. So it appears on the front page of Frontiers as soon as the paper is accepted and can be gathered on the specific sub-topics that will extend the exposition of the paper. So for instance, this research topic was co-sponsored by ICT and computer image analysis and no informatics in Frontiers in No Science. The two journals did the co-sponsor the special topics and when you submitted to the special topic, you said okay, my primary goal is ICT and my primary goal was Frontiers in No Science. There is also a commitment and also awarding and going to publication of specific conferences or to move workshops for instance to a special topic. The one I showed before was a workshop on Mikai Conference and it was submitted as special topics afterwards and so there is a big commitment to have a publication on conference through open access systems. We die with all types of references and on the major ranking of journals. So publishing in Frontiers has several advantages like fast publication in a field. Three months of publication is something which is not usual and it's important too and I think it's one of the very important aspects to have very fast publications but still peer reviewed and very serious peer review. It's open access. All papers are accessible by everyone. Copyright to author. Everything is transparent. I think it's very important this aspect that's pretty new. All comments and discussions between authors and reviewers are published and then you can see how much a reviewer did improve the quality of a paper and I think it's really a way to encourage and to grant the work of reviewers because now it's more and more difficult to have good reviewers for the paper or if a good reviewer make a good job then it appears that when a paper is published you can see that the reviewer has good review and good work. I think it's very important especially for promoting young reviewers in the film. And so it builds some research network and it turned it to have multidisciplinary topics, research topics promoted by this way. Okay, so I will move to the rest of my presentation. I don't know if you had few questions on this aspect before I move or I can move and can discuss afterwards. Okay, so now I will talk about imaging biomarker which is the film more where I'm working and especially an instruction on application on multiple theories. But many of the techniques I show can be applied to other types of brain pathology. So first, what is a biomarker? So biomarker is a feature that is objectively measured as indicators of normal or pathological biological process or changes of pharmaceutical response to a therapeutic intervention. So that means we have different types of biomarkers especially three major types which can be biochemical or historical parameter detected on tissues when you do a biopsy for instance that's that type of biomarker you are extracting. Biochemical parameters or cells obtained from liquids or blood or urine samples. It's typically when you go to a biochemistry sample test and finally, which is rather new, anatomical, functional, cellular or molecular parameter detected from imaging. But for this type of topic and of course I will talk only about these last categories of biomarkers, what are the major capabilities or requirement for an imaging biomarker so it can be used to detect pathology, to predict the risk and the level of risk of the pathology to classify the extent of the disease and to evaluate a therapeutic response. But they must be for that quantitative and not qualitative, accurate, reproducible and feasible over time. That means you can reproduce over time with similar results. So in order to provide imaging biomarkers, you need to address very large range, large scale of information going from imaging sensors to the patients through microscopy scale, mesoscopic scale, population scale, so it's several thousand of different metric scale to address or temporal scale to address. And that means you have to develop some specific quantitative imaging sequences to provide a not only qualitative sequences as we usually do. You have to work with some specific contrast agents to mark some specific markers in the tissues to analyze structures and functions and to derive some metrics on structures and functions at the population's case to work on the big data issue and the machine learning aspects on this population to build some or to infer some model and then to apply this model on patients in clinical medicine to see for instance how a patient is close or different to a specific group or a specific model learn on the population. So on MS, I will mostly talk about these three first aspects. And what is imaging biomarkers or what do we mean by imaging biomarkers in multiple cirrhosis and what is the goal? The goal is to guide the clinicians or the neurologists, could be a neurologist too, within the mass of information to integrate into the medical decision process. So multiple cirrhosis, I don't know if you're aware of that pathology, it's a chronic inflammatory and diminishing disease of the brain of the central nervous system. It leads to acute handicap in young adults and it's high prevalence in Norwegian, that's why we have a big clinic in MS and it's the most frequent CNS disease in young adults which leads to a disability and a high social individual cost. The main issue and challenges are to do the diagnosis as soon as possible for adapt the treatment of the pathology as soon as possible to prevent the disease progression and the future handicap to better understand the pathology because it's not very well understand here in order to have new in vivo classification of MS for instance and to set up the new therapeutic protocols, especially the new drugs which are arriving on the market like disease modifying drugs. And what is important in terms of epidemiology of this pathology, if you look at how the pathology in a population, you have a clinical score, a clinical score of three is when you start to have physical disabilities. And in fact, what you can observe is some people are reaching from T zero to a few months, this score three and some may take more than 20 years to reach this medium scale. Afterwards, we move at the same speed to a more severe form of the pathology and we don't really know why, in fact, some people are on the red line, some people are on the green line and how we can move people from one line to another line. So that's why most of the work and why we need imaging by your record is to better study the early stages of the pathology. So for that, the state of the art is, for instance, to detect the focal legion we have in the brain by using different types of mathematical procedures like for instance, we use an analytical model and we learn analytical model of normal appearance of the brain tissue and then the legion can be described as outliers, statistical outliers from the normal signatures of the brain tissue and by bringing some neurological expertise as rules to filter the detected voxels to be legions or not legion, then we have to embed some specific prior about the rules, what are the legions? So that means we can build some quite complex workflow for detecting from a set of images, usually it's multimodal images from MRI to be able to detect the MS legions. So this type of workflow goes from different aspects, denoting, inhomogeneic correction, registration, classification, for instance, by using Bayesian or robust expectation maximization solution, but these procedures is sensitive to presence of outliers, so we need to have robust estimations and then some kind of local classification to do region fusion, classifications and then for instance, we use graph cut in order to have a regularization solution based on local priors, we can build from this robust aim. And finally, we apply rules coming from the neurological signatures of legions in order to segment the legions. And at the end, that's the type of image you have at the beginning. With a pre-processing, you can clean the image to make the image better for detection. We can prove that the pre-processing improve a bit, improve not a bit, significantly the quality of the detection and then you can detect the legions and especially detect the extent of the legion which can be, which becomes an imaging biomarkers because the extension is one marker which can be used for evaluating the evolution and the stage of the disease. There are other type of techniques, we have developed other type of techniques, for instance, especially when we want to compare different scans over time, we need to have intensity normalization between two scans, but again, the intensity normalization shouldn't be sensitive to the presence and the evolution of the legion. If you have a standard intensity normalization, histogram normalization, then you will be biased by the presence of the pathology. So you have to define some robust intensity normalization, so you know we use a gamma-divan-john solution for intensity normalization in order to compare two different images our same contrast or two different imaging, when for one of the images we inject some contrast product in order to mark the inflammatory legions and so in that case, what we want to detect automatically is when a legion has been marked by this special contrast agent. And for instance, we are able to detect the evolution of the legion in patients between two time by having a difference and to have a priority of change in detections and to be able to show what are the new legions, what are the disappearing legions and maybe what are the false positives when we do evaluation. And that's another type of imaging biomarker which is even more useful than the extension of the legions we showed just before because what is really important is how many new legions and how many disappearing legions are evolving during the evolution of the pathology in order to grade the patient on the different lines. And as I said before, we can also do the same type of procedures between contrast and non-contrast images in order to automatically segment or give the extent or detect, just detect the legions which are getting the contrast. That means that the legions which are specifically acute because they are considered as inflammatory legions and we'll come to that afterwards why it's so important to detect these inflammatory legions. We have another type of tools which is more based on population models so we can make a population model of what is the normal appearance of the brain from MRI for instance. And when we do that, we can have a patient model which can be compared to group controls and by using statistical model of this group control, we can say that the different, the probability that some bulk source are different between the patient model to the group controls has a big chance to be an MS lesion. So that's the idea. And from that statistical detections, we can have some local organization. Then we can also detect some, after intensity normalization, what is probable abnormal brain tissue in the brain which are probable MS lesions. Here again, we have to apply the similar rules and I was talking before with a regular form. We can also address this issue with machine learning techniques. And like for instance, we use a technique which is called probabilistic one class SVM. When you do machine learning and the state of the art of the techniques, in fact, works well in practice when the training example are balanced. And the problem when you have a small multiple sclerosis lesion with a mass of bulk source coming from the normal brain tissue, you have a typically class imbalance problem which means that the lesions are represented within the whole set of other type of features. And this class imbalance should be introduced in the SVM classifier. So we have developed a specific procedure for the probabilistic one class imbalance program by using classically a training phase from multimodal images. We do just a dimension reservation for the features. And then we use a least square classifier in order to train the model. And with a train model, then we can compute a probability score of aggregation based on SVM which can be then threshold to do lesion detection. And then the testing phase consists in having a patch and having this patch scanning the images to assign some probability of imbalance and this probability imbalance. If you threshold this probability, then we can detect bulk source which are probably belonging to lesions. And when we apply that, we can compare from the input image the ground truth from an expert what we can get from a classical one class SVM and what we can get with the procedure when we have a probabilistic detections with one class learning imbalance SVM. Of course at each stage, we need to have evaluation of these procedures. Another aspect, also a machine learning strategy in that case is to use dictionary learning and sparse representation to detect the lesions. And again, we can consider the atoms describing the pixels of the lesions are abnormal signatures so if you're not aware about sparse representation, so sparse representation used to represent signal using a linear combination of few basis functions which we call dictionary and sparse vector. So sparse representation can be seen as an optimization problem where we want to minimize this probabilistic expression with respect to a penalty term alpha which is a sparsity penalty term and usually we don't use the L0 norm because it's more difficult to optimize with rather use the L1 norm. And the related dictionary learning strategy is when D is unknown and in fact we try to find D such that each signal can be represented by the sparse linear combination of its atom and atom is one component of the dictionary and so we have to minimize this expression according to each class of data we want to classify with the dictionary. So for instance, if I take the image before, for instance you can use one class for the normal tissue and one class for the lesion, you have the class imbalance program or you can use one class for the gray matter, one class for the white matter, one class for the liquid and then one class for the lesion when the class imbalance is not that strong but then you have to learn more information. And afterwards the classification is just to find the K class or two or four, for instance, which minimize this expression. So in practice, what we discovered is the size of the dictionary is exactly, behave exactly as the class imbalance problem in SVM. That means if you as, usually when you do dictionary learning you don't change the class, the dimensions of the dictionary according to the complexity of the tissue you want to classify. And in fact, if you apply this, you don't get exactly, you don't take at all the same type of weather. For instance, here you should use two dictionaries with 5,000 atoms, one for dictionary for healthy class, one for the lesions. The red are true positive and the blue are false positive. So you are increasing the number of positive which is not good for the detection. And if you adapt the ratio of the dictionary with two big difference ratio, 5,000 for the normal tissue and 1,000 for the lesion class, then you have false negative classification. So that means you have to learn the right number of ratio between the different dictionary classes. And here for instance, if you have two classes between four dictionary with 2,000 atoms for the white matter, gray matter and CSF and 1,000 for the lesion class here, which is the best results we can obtain on that type of image. So the problem of class imbalance that was describing for SVM is exactly the same for sparse representation using dictionary learning. And we did not test for instance deep learning aspects but the relation between dictionary learning and deep learning is very similar because in fact if you want to do deep learning from that aspect, you just have to learn different hierarchy of dictionary to have similar effects and the multi-layer deep learning aspect. But then once you do that, you're happy because you are able to segment the religion but in fact what we discover is it's not enough because multiple sciences is more complicated by just segmenting the brain lesion and there is what we call a radiological clinical paradox which means that if you have a quantification of the number of lesions, the extension of the lesion, the evolution of the lesion, still you have a difference between this type of markers, quantitative markers and what the clinics provide, the sign of the clinics and that's because we are missing information. For instance, the lesion severity, the position and the macrostructure of the brain. And for that we need to change the paradigm in order to go beyond this famous clinical radiological paradox in order to measure what the human high cannot see on images. So far we were just detecting what the radiologist was able to see but we also needed to go beyond in order to study, for instance, the inflammatory process. What is behaving in the cells, in the inflammatory cells, especially the macrophages or the microglia which are inflammatory cells in the brain. We are also, we need to study the characterization of axonal degeneration in the brain to better quantify how the axonal degeneration is occurring. And for that, we had conducted two different studies to study the inflammatory cells and the macrostructures by using, in the first case, nanocurriers of iron oxide which can tag the macrophages activity in MRI by using this specific contrast agent. So that's a very new contrast agent. In fact, it has been discontinued because it's not profitable for the pharma company but I can show how much, in fact, this new contrast agent can be very important for better characterization of the disease. So here again, what we did is we studied the longitudinal analysis of only the inflammatory lesions of the one with this new contrast agent and the regular one, the gadolinium one I was talking about before. And then the goal is to discover the different classes of lesions that can prospectively stratify the multiple sclerosis population. So we take the, we take the different time point and in fact, we took only the first two time points of the lesions and with the first two time points we are just doing two layer classifier. The first one, so we use a spec-prostering classifier because we don't know in advance how many classes we want to find. And just by using shape information of this lesion, inflammatory lesion, so a very simple feature space, we are first able to classify the different types of lesion we can observe on the different image in the population. We are just a group of 35 patients with two time points for all these 35 patients. And then according to this first layer classification, how many classes of lesion we can recognize, then we are able to see for each of the patients how many lesions of each classes this patient can have. And by doing another classification using spec-prostering, we are able, afterwards to classify the severity and the group of patients which are most probably will have a severe evolution of the pathology because we are comparing this with other type of markers we have discovered two years afterwards. So that means we are able with this technique, in fact, to have early characterization of lesion patterns in the brain at the very beginning of the pathology because it was the first scan when they arrived to the hospital. And with the two first scans, we are able to predict what will be the evolution of the patients two years afterwards. So it's because we had addressing issue of what the human high cannot see with inflammatory lesions. Another aspect is also to study the macrostructure. So maybe you're more aware of these technologies, typically how to better use diffusion imaging and to go beyond the current limit of diffusion imaging in the brain and in order to extract different parameters of macrostructures in the brain by using diffusion MRI, going beyond DTI, using multi-compartment diffusion MRI and using also multi-compartment models of the macrostructure. So diffusion MRI, for instance, in the brain, if you do tractography, you can reach different parts of the brain with a classical tensor model. But when you use a multi-compartment model of diffusion, then you have a more specific and sensitive way to address a different connection in the brain. And we can validate using fMRI activation where we have better capability to connect the different true connected regions in the brain by using this type of process. And in addition, we can then characterize how the fibers in the brain are crossing and how we can compute multi-compartment of these different fibers crossing together. So that's how to use diffusion imaging and we are able to better characterize MS lesion with that aspect or to just do some tractography. And another aspect is to use also quantitative MRI, which is called MR relaxometry, especially the T2 decay of the relaxometry of the tissue. And when you look at more carefully the T2 decay, you can see three different compartments, one which is specific to the myelin, one is specific to the axonal cells and what is specific to the water, the free water in the brain. And the goal is to estimate the three different compartments and so we can do some estimation on this specific sequence in order to extract free freed weights, water-mailing fraction, so the fraction of myelin in the brain, especially when you have an MS lesion here, for instance, then you can see that the water-mailing fraction is decreasing, which is another quantitative marker of what is occurring in the brain. And the axonal cells, for instance, is still doing the same in that case. But the interest is also to look at this technology a long time to see how, for instance, tissue, which are pathological tissue, are evolving on the time. And for instance, between three different time points here, we can see that the myelin reduced over time within the lesions, or the axonal cells is also reducing over time according to the lesions, while the water accumulation in lesion is increasing over time. So here you have a kind of new descriptors of evolution of the tissue, the macrostructures within the brain lesions that can be characterized by the estimation of these different compartments from the quantitative MR sequence. So I was talking about what can be considered biomarkers from images, especially quantitative biomarkers from images on the specific pathology. And now I will talk about how we can do population imaging and treat patients from population imaging at a larger scale by using digital infrastructures. So we'll just review a few points about what in medical imaging, what are big data applications and why we are interested in this domain. So for instance, in medical imaging, medical image database are increasing between 20 to 40% every year, and a regular hospital is providing, for instance, up to 300 terabytes of images every year. In the USA, for instance, because the statistics have got was from the United States, 30% of imaging actions are produced for legal protection. That means we produce images not for the treatment or for the diagnosis, but just for legal protections. And they are part of 400 million medical actions per year. And it's probably the same in Europe or even more. And 1.3 billion of the medical images were stored in 2010. That means more than five billions of images are stored now in Europe or in the US. And this increases the need also, especially for legal protection of long-term storage from today, so already 20 years, but can go to the lifetime for the people, so 70 years, for instance, at the Euro recommendation. It involves also big aspects on data analytics for image processing and machine learnings. It involves aspects on data protection with complexity or constraint. And it's also very different to the e-commerce domain when the big gap, for instance, are addressing this issue. And they think that they are, they can move from big data e-commerce to the medical imaging domain. But the problem is different because you have less instances, you have more data per people, for instance. Much more complex data per people. So we need to adapt the computational solution for managing this problem in the medical domain, especially for medical imaging. So where we are right now, so what do we do with the data produced? So we have an explosion of production and exchange solutions for imaging data with internet, but information doesn't mean knowledge. It's not because you gather images that you gather some knowledge on the images. You need to extract information from images. You have, you need to extract the metadata of the images or to represent the metadata. So they are no knee-jewel about how to exploit this mass of information easily and to deal with this mass of images. And for the moment, they approach our rather base on the descriptive analysis than on statistical one to search for correlation, but correlation doesn't mean inference. And the idea is that the mass compensates the quality which is not true. It's not because you have massive type of information. If the quality of the information is not good, then you don't have the quality you need to reach. So we go toward a generalization of digital infrastructures, what we call a PAX 3.0, with where local storages are overpassed. We need to have storages on the cloud. We need to have this association between acquisition and storage and to have remote viewing and analytics solution. We go to this way by the emergence of dedicated digital infrastructures, but the problem is what type of operators to run this infrastructure, what cost and cost model to use these infrastructures, how to promote emergence of virtual communities of users and the emergence of new e-services on top of these infrastructures and the images hosted by the infrastructures, and the emergence of new usages with new way of working. Now the image is shared, is not locally used by the local radiologists. We have emergence of virtual care networks. We have evolutions of the concept of territorial coverage toward, for instance, specialized imaging centers on the territory. And we are also emergent of image sharing and processing through the standardization of the imaging protocol. That's an important aspect that I will, it is like this afterwards. And we can have very huge potential economy of scale by using these infrastructures. And why we need also for research to go to open data because we need to access more and more database for building and classes, learning models, data mining and such of similar case for education, for training the professional, for evaluation or validation of data analysis, image processing solutions, for certification of digital solutions around the images. And they are encouragement from funding agency to the open data models with the emergence of the population imaging concept. So they're still moving toward open data, some major issues, so for instance, what types of operator we want to run these infrastructures are their public, private, national, international, global, like GAFAM for instance, which are targeting this aspect, which economic model who bear the cost, who will pay for these solutions, what standards for exchanging, for interoperability between infrastructures, how to perform data quality control because again the mass doesn't compensate the quality and how to promote the emergence of new players like network operators, data centers, startups, but also big companies like GAFAM which are addressing this domain. And we need also to evolve or to adapt the regulation to these new domains and at the end we have also to address the ethical issues. So medical imaging can be seen as a service but with specific requirement because the data are heterogeneous, multi-stage, they have a strong semantic, we need to store and to organize the semantic, we need anti-acquisition protocol normalization, standardization of the acquisition protocol because now more and more the protocol are distributed over sites in order to make some networks of clinical users and the data are also confidential, so we have security issues and we need long-term sustainability of the data. According to the data analytics we met processing solution, they are often correlated. We use workflow, I describe some very complex workflow and it's more and more the case. So we need automations of this workflow on large cohorts. We need quality assessments for transforming the image from qualitative to quantitative information and to provide some reference values what we call imaging biomarkers and I'll show you what is imaging biomarkers in the context of MS. The computation time can be high on the population cost so you need big computational infrastructures but it can be also critical, for instance when you do real-time simulation or you have to perform image access or processing in an emergency room. So we go more and more where a federation of data and services and we'll just rapidly destrate solutions to provide the solutions in this context of medical imaging as a service which is a Chanois platform we have developed in Rennes and which is installed in different infrastructures. So the goal of Chanois, it's a software as a service environment to manage population imaging data because clinical research involves more and more and mostly now multi-center studies. So we need data storage and archive solution. We need data structurization for data sharing and accessing over the internet. So existing solution like CD, packs are not at all available for that type of solution. So we are providing this new solution which is a secure web server accessible for any web browser and it's currently operating on more than 50 centers of clinical research with more than 200 user connected. Austin, probably around 100 studies right now, different clinical research study with more than 3,500 subjects connected. So it's a web portal with different solutions and online visualization, anonymization or identification of the data. Download it or store data. It supports process data. So if you process the data, you can re-import the systems. Of course, we have import solutions for CD-ROM, packs or standard Nifty format. We support clinical scores for instance like no psychological scores. We have also, of course, user access control using a single sign on solution. And we support multi-center research studies. And we have implemented this solution for the national court in the meeting post-leurosis which is called the OFFSEP court. So the OFFSEP court has a goal to gather about 40% of the French population in MS. That's something like 50,000 patients at term. And so the challenge for us was to first do protocol standardization for the MRI in order to have a minimum standardization everywhere in France for small location and big locations of MR. We had to be compatible with a very large set of MR manufacturers and MR scanners. And the same for packs system because local hospital of the images are stored on local packs and need to access to these local packs. This system is also right now implemented in virtual care network in Western France where all patients in MS will go to this process. Whatever the place when they have their MR scanner. So they use the previous protocols of standardization of MR for MS. And then afterwards, there are kind of complex system because it's really for clinical care over the internet. So we have a security aspect and Shenouar will be used as a hosting solution for the data set. And finally, I'm managing this infrastructure in France. We are developing a national infrastructure in France which is called France Life Imaging and especially the information and I am information analysis and management node which are the goal in fact to provide at the national scale computational infrastructure for medical imaging not only for no imaging where a user can be a clinician so can be a computer scientist researcher can access to the web portal and this web portal will have data repository solution and Shenouar is one of the data repository solution and we are able to have interoperability between different data repository by using metadata adaptations using solar catalog and we have a computational infrastructure which are able with some app store in fact to provide analytics solutions which can be run on a local system by using for instance made in regards of visualization and computational solution or brain visa or by using a workflow management which is the V-platform which can run then different workflow on mass of data on HPC solution like EGI grid for instance and of course the derived data can be also re-import in the distributed archive and in fact we had a project together with France Life Imaging and the offset on MS infrastructure on MS I showed in France to organize this year a MIKI challenge for validation of the quality of multiple sclerosis lesion segmentations which was developed by using these infrastructures and in fact the data were hosted on the, no this wasn't here the data were hosted on Shenouar on the data repository and with all workflow of image processing were run by the V-platforms on the dedicated clusters in order for the challenger in fact to not run themselves their workflow but to have the platforms to run for them the different data of the competition. So we are under the way to make this infrastructure public and available to everyone by providing the different layer of the cloud the cloud solution providing just an infrastructure as a service solution or a platform as a service solution or at the end server as a service solution with a specific service provider for the end user we have analyzed a different scenario of usage of this infrastructure but I won't go to the detail because I'm running out of time and so I will just finish my presentation so how archiving and sharing images will change the usage because gathering of the data remains an issue it requires reconciliation of image sources reconciliation of images with clinical context do image fusion quality control harmonization of the protocol we know we need to do that but it's not easy to do it and the community need to practice these different new usage image analysis remains an emerging field we need to test the robustness of the tool to limit the computational time the raw producibility of results that's why computational infrastructures are very important in order to see how much and open data the results are reproducible from one system to another system and that's clearly a big evolution for the next decade who wants to share their data there are still some conservatism of the community but less and less more and more now in clinical research the PI agree to share their data as soon as they have their own advantage for instance citations or results of the image processing you can provide on the data but of course we are also to face some regulatory constraints and we have to make these regulations evolutions and to be partner of this evolution who can share their data so we need to go beyond a club of insider because many clinicians involved in clinical studies are not insiders on technologies but they want to be part of this club because they want to work in a virtual network of users and it's really a move of what we will bring in terms of sociology of clinical research this type of new digital solutions we need to offer customized solutions according to the different types of user to have a kind of certification of the provided solution for instance you provide an image processing solution on an infrastructure how you can ensure that the solution which is provided is certified so having a kind of validation on test data on open data is a way to promote the quality of a data analytics solution but we are also to accept a fair cost for this new use if we want to use if we want to disseminate or technology or to disseminate or data to work in a virtual network there is a cost and who will pay the cost probably we will have to pay the cost on all specific ones for instance how to ethically manage data sharing and open data so we need to anticipate before the programs arise collectively through legal revolution of the legal and individually by even now inserting in the information constant on the patients that the data will be provided for research for instance or will go to open data solutions and it's very important to address this issue as soon as possible security is and will always remain a challenge this is true especially for the digital domain we know that security of digital solution is acute and it will never be solved because it will be always attacked on the system so we should never underestimate this issue we should never excuse ourselves to not taking care of security issue we also have to be careful that the the level of the response to this risk doesn't kill the use that means if you make a level of response too high in terms of legal aspects or security aspects no one will use this digital solution so we need to adapt the risk the level of risk at least we need to adapt the response to the risk to the level of the risk and it's a multidisciplinary aspects IT security is just it's not just a medical issue it's an issue in digital science at large and the medical domain needs to work with ICT people in these aspects and so we need also to invent new jobs and develop these new jobs because nothing can be done without a strong integration between engineers, researchers, lawyers and doctors in that field and that's it thank you for your attention it's a good question what I'm putting here in fact it's not a it's decentralized here it's clearly decentralized because in the in the from self-imaging infrastructure I'm waiting for that in fact we have three different archive solutions and the archive solutions are located at different places so what we promote is more the interoperability between the archive solutions when you go to the portal the portal is a web portal but it's more web portal to orient to offer an interface for the user but the infrastructure itself is totally deported on the internet so for instance when you access the web portal for the data and you may have some access controls on some of the repository and then you will be able to by query the data to see where the type of information you are searching are hosted or not hosted on the repository and only afterwards you will get some solutions and then you will be able to access if you have wanted access to the actual repository of the data definitely in medical imaging what you say is more true for biological imaging but in fact this type of infrastructure I always talk about not clinics but clinical research and I think there is a big difference in clinics you may have right but there is no way to share all data on clinics there is no interest in that but what is interesting is the quality of the data you provide in clinical research and in that case the data can be it's more research infrastructure it's not for clinics what do you mean by radio mix to be sure it's another way to talk about machine learning on images you can do radio mix on that type of systems for me it's machine learning in order to train some classifier to detect or to represent some signatures on images for instance there are a new interesting fingerprinting solution in EMR for quantitative EMR when you have only one acquisition and then you can simulate all types of contrast on EMR and to do that you need to learn how the data base how the contrasts are displayed in order to put the right the right parameters in fact for balancing the different quantitative EMR acquisition so it's typically based on the technology you are talking about but it's done in fact it's done on the computational infrastructure no more it's a good question of course for that type of problems we are also thinking about how now we can for instance cross correlate the type of feature we extract from quantitative imaging with genetics for instance that's how it would be very interesting to interpret to have additional information to better type of data we extract from here and in terms of these frustrations we are also working on this aspect because in fact we are discussing with people who are running bioinformatics infrastructures which are doing the same thing that what I showed in Franislav imaging but on biological imaging which is a different issue in biological imaging too because the data are really big and are more difficult to move so they have some different issues and so now we are thinking about how we can cross information from biological imaging medical imaging and genetics and at which semantic level probably more semantic levels than raw digital levels where pixel cannot be compared with the speed because it's not really meaningful and so that means these infrastructures for me the right way is not to do one single infrastructure which do everything but it's more to have inter-appearable infrastructures that means medical imaging infrastructures the one we are doing is able to talk to bioinformatics infrastructures to extract the relevant metadata information in order to have a joint analysis of data coming from the same context so it's more that way we are it's a good question also right now the regulation the ethical aspects is not adapted for this new so there are some normal regulation for providing the data but right now the notion of how the data will be hosted shared are they going to open data or not open data what type of information you can in a clinical research protocol for instance you may put some data in the open domain but some of the data are too confidential to be put to the open domain and these aspects are not recovered and the risk right now is to adapt the level of regulation too high and if you put the level of regulation too high then you you limit the use of this new technology so that's really a risk so I think that's type of message we have in France to the regulatory people that we need to discuss about how the regulation should be adapted to this new evolution yeah one aspect in fact in our infrastructure for instance in the infrastructure we have an app store and in the app store we have also some workflow managers so what you're talking about is how a community of people can build together even if they are not in the same place a kind of workflow set of tools comes from the app store and are organized together as a workflow and then you can have adopt the same language for the workflow in order to build a pipeline which can be run on the infrastructure so right now we have a tool for building a simple workflow otherwise there are tools in the VIP platforms to run the workflow but we are not distributing this system yet that could be interesting for us the principal aspect is to have kind of app store when people can store their solutions which can be run either on standalone solution on a remote solution thank you