Bridging the gap towards efficient Collaborative Digital Pathology: a pioneer initiative of the FlexMIm project D. Racoceanu1, P. Bertheau2, A. Veillard3, D. Ameisen4, B. Ben Cheikh1, N.Phan2, F. Capron5, C. Guettier6, P. Brezillon7, J.B. Yunes4, J.F. Pomerol8, V. Verger8, J. Klossa8, P. Constant9, F. Lallemand9, B. Girard10, L. Toubiana10, J.M. Temerson10 1

Sorbonne Universités, UPMC Univ Paris 06, CNRS, INSERM, (LIB), Faculté de Médecine Pierre et Marie Curie, 91 bd. de l’hôpital, 75013, Paris, France, [email protected] 2 Service de Pathologie, Hôpital Saint Louis, 1 avenue Claude Vellefaux, 75010 Paris, France, [email protected] 3 Sorbonne Universités, UPMC Univ Paris 06, CNRS, (IPAL), 1 Fusionopolis Way, #21-01 Connexis, Singapore 138632, Rep. of Singapore, [email protected] 4 Université Paris Diderot, CNRS, (LIAFA), 8 Place Aurélie Nemours, 75013 Paris, France, [email protected] 5 Sorbonne Universités, UPMC Univ Paris 06, APHP, UH Pitié-SalpêtrièreCFx, Dpt of Pathology, 75013, Paris, France, [email protected] 6 Université Paris Sud, Hôpital Kremlin-Bicetre-APHP, 78 Rue du Général Leclerc, 94270 Le Kremlin-Bicêtre, France, [email protected] 7 Sorbonne Universités, UPMC Univ Paris 06, CNRS, (LIP6), 4 Place Jussieu, 75005 Paris, France, [email protected] 8 TRIBVN, 39 rue Louveau, 92320 Chatillon, France, , [email protected] 9 Pertimm, 51 Boulevard Voltaire, 92600 Asnières-Sur-Seine, France, [email protected] 10 Orange, 28 chemin du vieux chêne, 38243 Meylan, France, [email protected] Introduction This paper presents the main results of the FlexMIm cooperative project launched in January 2013 by a consortium of six partners in the area of Digital Pathology. The project architecture is relying on secure cloud hosting of medical data (whole slide images being uploaded on to an ad-hoc secure image server). The original target was to develop and provide, within a cloud, a complete set of innovative digital algorithms for Whole Slide Image analysis, annotation and retrieval, as to develop Semantics algorithms for high-content image management. We describe these smart

algorithms enabling and supporting an extensive collaborative telepathology initiative intended to address a large community of pathologists, mainly coming from 27 pathology laboratories in the Paris region (coordinated by the public hospital system of the city of Paris and its suburbs – APHP: Assistance Publique - Hôpitaux de Paris). A first glance of the assessment by the pathologists is given. Project organization and consortium At the very early stage of the project, one major event occurred which slightly modified the initial goals. ARSIDF, a second project launched by APHP with the involvement of a subset of FlexMIm consortium, was labeled and funded in the field of Telepathology, for a collaborative pathologists’ work on common medical cases. This gave the opportunity to specialize both projects on complementary goals. Therefore, ARSIDF focused on pathologists’ collaborative work including web conferences sessions, secure hosting of digital medical images in the case of image sharing and remote diagnosis for 2nd opinion in difficult cases. Accordingly, the FlexMIm project became a pure research project, aiming at improving the global workflow of digital pathology and targeting development and use of the necessary smart digital algorithms facilitating pathologists’ work. Orange led the project management (WP0) and the associated communication (WP7). WP1, led by APHP, was dedicated to the constitution and coordination of the cluster of pathologists’ working groups, and to the preparation of a full information document that would tailor the platform specifications in accordance with the pathologists’ requirements. WP2, led by Orange, consisted in administration of a hosting platform in cloud mode, and the development of an innovative algorithm for image best transport in situations where network conditions could be subcritical. WP3, led by the UPMC University Paris 6, was dealing with semantics and images and involved Pertimm, a company specialized in data indexing and retrieval. WP4 was driven by TRIBVN (a medical software company with large expertise in whole slide imaging), which developed dedicated image analysis algorithms, and integrated on the hosting platform all of the algorithms developed by the partners, including (but not limited to) the blur detection algorithm developed by Paris 7 University. WP5, dealing with coworking activities, was actually transferred in the ARSIDF project, as previously explained. WP6, led by APHP, consisted in permanent advice during the algorithms development phase, by including the current (six month - started in December 2015) phase of full assessment of all the algorithms produced by partners (and the potential upgrade requests by the partners after this evaluation) during the project.

Algorithms and achievements This section gives an overview of the algorithms developed by FlexMIm partners and integrated in the Teleslide-FlexMIm platforms since the beginning of this project. The details, goals and the operation modes are listed in accordance with a possible query in the medical process workflow. In order to avoid uploading unusable slides, the platform is able to filter potential flare or scanning defects in the slides. During the FlexMIm initiative, this was achieved by the development of a robust blur detection algorithm by University Paris 7 [5]. The algorithm was first tested in standalone mode and found successful with nearly any slide file format (JPEG, PNG, TIFF, SVS, NDPI and MRXS) prior to its integration in the project platform. Once integrated, optimizations were made in order to remain compatible with most of the existing slide file format. The optimization of the pathologists’ access to WSI using their terminals, by providing this access as a natural function of their actual environment (defined by pathologist localization on the network, the network bandwidth, the type of terminal used, the medical use of the image e.g. diagnosis or training), is insured by a “smart transport” algorithm, developed by Orange. Tested first in standalone mode and then integrated on the FlexMIm platform, it allows the pathologist to define a quality and a time thresholds, for an efficient image access, in accordance with these thresholds. An important improvement of displaying speed has been shown, as much as 10 to 50 times faster in poor bandwidth context, while the displaying speed was almost identical with or without this smart transport algorithm in good bandwidth context. In order to deal with possible staining variations (mainly due to interhospitals variations in the dye chemicals used for slide preparation), a powerful coloration correction algorithm was developed by TRIBVN. In the field of automatic detection for some specific cases, an analysis algorithm allowing help for identification of possibly critical (dysplastic) / uncritical areas was developed for Inflammatory Bowel Diseases (IBD) by TRIBVN and tested. It was further integrated to CaloPix Server and then made available on the collaborative platform. Virtual slides can be annotated on line due to several value-added functions developed by TRIBVN on the CaloPix server. Besides, Ki 67 index marker being essential in Breast cancers or in neuroendocrine tumors, a specific algorithm allows counting stained and unstained nuclei. Indeed, the ratio of stained/stained+unstained cells (Ki67 index) represents an important criterion for assessing the prognosis in tumors. For both IBD cases and prostate cancer cases, a contextual graph method has been elaborated and implemented with the support of UPMC Univ.

Paris 06, and the derived vocabulary put online on the project platform. Besides, mitosis detection algorithm has been built and tested. This H&EWSI based algorithm can be launched off-line for specific localized second opinion requests. In the same spirit of the above described tools - designed to provide quantification support to pathologists, on line semantics service / vocabulary has been produced to annotate IBD slides, with the ontology repository built by UPMC Univ. Paris 6 and an external search engine developed by Pertimm. The whole platform is linked through APIs used by the Pertimm’s search engine to communicate with this ontology, and to grab and index annotations. The ontology has an editing interface designed to allow pathologists feeding it with data, that will enrich it and improve the annotation, indexing and retrieval tools, as making possible converging towards a strong semantics tools open to everyone. Moreover, a moderating mechanism is being implemented to improve the overall acceptability of the ontology updates by all pathologists of the project. Algorithms assessment and methodological approach In order to assess the performance of the above-described algorithms, a complete system of on-line forms has been implemented by TRIBVN, upon recommendations of project pathologists in the “build” phase. During the evaluation pathologists “users” can thus provide quantitative and qualitative answers to specific questions without interruption of their tests and no risk to lose their written answers. A statistical treatment of these answers can then be derived, provided that enough pathologists are enrolled in the evaluation. All algorithms evaluations are performed on usual PC but pathologists were also provided with some touchscreen tablets to assess its ease of use in Digital Pathology. For efficiency reasons, in order to reach a minimal consistent cohort from the 27 pathology laboratories in the Paris region involved in the project, a focused number of laboratories were equipped with a new-dedicated slide scanner. For these services, an engineer is fully assigned to the support of the pathologists of these laboratories during the “build” and “run” phases of the assessment. First results, conclusions and follow-up In view of the very first results coming out from the beginning of the evaluation phase, we can conclude that most algorithms, which have been tested by the “FlexMIm” cohort of pathologists, have great potential for the development of Digital-Pathology and Telepathology. However, the most interesting resides in the fact that the complete combination of digital tools,

methods, and algorithms creates a genuine arsenal of added value functions, which will support the pathologists in their medical practice, in a very near future. Acknowledgment Authors and FlexMIm partners acknowledge the support from the following Funding partners: BPIFrance, Regional Council ÎLE DE FRANCE, Department of Paris. References [1] [2] [3] [4] [5]

Ameisen D., et al. (2014). FlexMIm: towards efficient/effective Collaborative Digital Pathology, European Congress on Digital Pathology, Paris. Wootton R., Craig J. (1999). Introduction to Telemedicine, Eds., London, Royal Society of Medicine Press, pp.68-73 Racoceanu, D., Capron, F. (2016). Semantic Integrative Digital Pathology. Insights on micro-semiological semantics and image analysis scalability, Pathobiology, Special issue on Digital Pathology. Racoceanu, D., Capron, F., (2015). Towards Semantic-Driven High-Content Image Analysis. An Operational Instantiation for Mitosis Detection in Digital Histopathology, CMIG - Computerized Medical Imaging and Graphics, Elsevier, vol. 2, pp. 2-15. Ameisen D et al. Towards better digital pathology workflows: programming libraries for high-speed sharpness assessment of Whole Slide Images. Diagnostic Pathology. 2014 Dec 19;9(Suppl 1):S3. Professor at Sorbonne Universities, University Pierre and Marie Curie (UPMC Univ. Paris 06), Daniel Racoceanu is co-leading a research team dedicated to cancer diagnosis and therapies, at the Biomedical Image Laboratory, French National Center for Scientific Research (CNRS), French National Institute of Health and Medical Research (INSERM), hosted by the Faculty of Medicine Pierre and Marie Curie, Pitié-Salpêtrière, Paris. His area of expertise concerns biomedical engineering and imaging, being involved in a successful series of national (ANR), European (EIT Health) and international (A*STAR Singapore, NIH etc.) projects. Philippe Bertheau is Professor of Pathology in University Paris Diderot and Hospital Saint-Louis in Paris. He is specialized in breast and GI tract pathology. Since 2005, he has been involved in the development of a virtual slides website for university students and researchers in University Paris Diderot. More recently, he joined two digital pathology projects in the Paris area, one devoted to diagnostic telepathology and one (FlexMIm) more centered on technological and semantics improvements using virtual slides databases. Antoine Veillard is a researcher in artificial intelligence. His current topic of interest is deep learning. Antoine graduated form Ecole Polytechnique and obtained a Ph.D. in computer sciences from the National University of Singapore. He was attached to Université Pierre et Marie Curie for the duration of the FLexMIm project.

David Ameisen is a PhD in bioinformatics and a computer scientist. Member of the scientific committee of the European Congress on Digital Pathology, his research interests are communication standards and interoperability issues, gigapixel image quality assessment and digital pathology. Since 2007, he is the system administrator of University Paris Diderot’s digital pathology platform. Since 2015, he is working with the French public incubator Paris Biotech Santé to launch his startup on automated image quality assurance. Bassem Ben Cheikh is PhD student in biomedical image analysis at Sorbonne Universities, University Pierre and Marie Curie (UPMC Univ. Paris 06). He is working on new tools dedicated to tumor architectures and tumor heterogeneity characterization from histopathological images. He is specialized in mathematical modeling in biomedical imaging, image segmentation and classification. Nicolas Phan is an engineer specialized in health science, especially biophysics. He’s working at Saint-Louis Hospital as a research engineer for the FlexMIm project. His role has been to test the various algorithms, as well as supporting the pathologists in the platform’s use and algorithms’ evaluation. Prof. M.D. Frédérique Capron, head of the Pathology Dpt. - Pathology Imaging & Modeling Unit UIMAcP of the Hospital Pitié-Salpêtrière Paris, has been opening the way to the digital pathology protocols in France. She is the president of ADICAP (Association for the Development of Informatics in Cytology and Pathology). Prof. M.D. Catherine Guettier is Professor of Pathology in University Paris Sud. Former President of the SFP - the French Society of Pathology – she is now in charge of the diagnostic telepathology network in the Paris area. Patrick Brézillon is Full Professor in Computer Sciences at UPMC. His research domain is Context Modelling and Management in real-world scenarios using the Contextual-Graphs software. He has 460 papers in reviewed places and organizes the conference on Modeling and Using Context (CONTEXT-17) in Paris.

Jean-Baptiste Yunès is associate Professor in the IT department in University Paris Diderot (LIAFA) and specialized in networks.

Jean-François Pomerol is the managing Director of TRIBVN. He is in charge of developing the company towards new solutions in digital imaging. Graduated of HEC business school, he began his career in audit and finance, notably at Societe Generale, before joining TRIBVN in 1999.

Vincent Verger is a medical information system consultant. At TRIBVN he is in charge of the management of all digital imaging projects, for Universities, Hospitals, private laboratories or pharmaceutical industry. Jacques Klossa is the President and Founder of TRIBVN. He is in charge of R & D programs, of their implementation and of the coordination of research consortiums and academic partnerships. With an academic training, graduated in Industrial Marketing from "Arts et Metiers" engineer school, he spent most of his career in multimedia and image business. He is co-founder of IHE Europe Pathology and has been the Co-Chair of Working Group 26 of the DICOM Committee (digital pathology), from its creation in 2005 to 2012. Patrick Constant is chairman and CEO of Pertimm, an information retrieval company founded in 1997. He is a Chartered Engineer from Telecom Paristech, Ph.D. Computer Sciences and Linguistics. Patrick has held several positions of Natural Language Professor. He is also member od the Knowledge Thematic Commission of the Business Cluster Cap Digital. Frédéric Lallemand is an R&D engineer working like a front-end web developer at Pertimm. After Biopool European project about online biobank, he is involved in FlexMIm project with the aim of setting search engine for the platform of this project. Bruno Girard is R&D engineer at Orange Labs working in the “Health & Data” project. He has been involved in transportation optimization problem.

Laurence Toubiana is the Project Management director of FlexMIm. Marketing manager at Orange Healthcare for the product line around Medical Imaging, she is involved in many projects about Telemedicine.

Jean-Marc Temerson is R&D engineer working in “Health & Data” at Orange labs. After many collaborative projects management and preparation, he is now deeply involved in the Flexible Medical Imaging project targetting Digital Pathology.