La gradation de lames de biopsie fournit des informations pronostiques essentielles pour le diagnostic et le traitement. La détection et le comptage manuel des mitoses est un travail fastidieux, sujet à des variations inter-et intra- observateur considérables. L'objectif principal de cette thèse de doctorat est le développement d'un système capable de fournir une détection des mitoses sur des images provenant de différents types de scanners rapides automatiques, ainsi que d'un microscope multispectral. L'évaluation des différents systèmes proposés est effectuée dans le cadre du projet MICO (MIcroscopie COgnitive, projet ANR TecSan piloté par notre équipe). Dans ce contexte, les systèmes proposés ont été testés sur les données du benchmark MITOS. En ce qui concerne les images couleur, notre système s'est ainsi classé en deuxième position de ce concours international, selon la valeur du critère F-mesure. Par ailleurs, notre système de détection de mitoses sur images multispectrales surpasse largement les meilleurs résultats obtenus durant le concours. / Digital pathology represents one of the major and challenging evolutions in modernmedicine. Pathological exams constitute not only the gold standard in most of medicalprotocols, but also play a critical and legal role in the diagnosis process. Diagnosing adisease after manually analyzing numerous biopsy slides represents a labor-intensive workfor pathologists. Thanks to the recent advances in digital histopathology, the recognitionof histological tissue patterns in a high-content Whole Slide Image (WSI) has the potentialto provide valuable assistance to the pathologist in his daily practice. Histopathologicalclassification and grading of biopsy samples provide valuable prognostic information thatcould be used for diagnosis and treatment support. Nottingham grading system is thestandard for breast cancer grading. It combines three criteria, namely tubule formation(also referenced as glandular architecture), nuclear atypia and mitosis count. Manualdetection and counting of mitosis is tedious and subject to considerable inter- and intrareadervariations. The main goal of this dissertation is the development of a framework ableto provide detection of mitosis on different types of scanners and multispectral microscope.The main contributions of this work are eight fold. First, we present a comprehensivereview on state-of-the-art methodologies in nuclei detection, segmentation and classificationrestricted to two widely available types of image modalities: H&E (HematoxylinEosin) and IHC (Immunohistochemical). Second, we analyse the statistical and morphologicalinformation concerning mitotic cells on different color channels of various colormodels that improve the mitosis detection in color datasets (Aperio and Hamamatsu scanners).Third, we study oversampling methods to increase the number of instances of theminority class (mitosis) by interpolating between several minority class examples that lietogether, which make classification more robust. Fourth, we propose three different methodsfor spectral bands selection including relative spectral absorption of different tissuecomponents, spectral absorption of H&E stains and mRMR (minimum Redundancy MaximumRelevance) technique. Fifth, we compute multispectral spatial features containingpixel, texture and morphological information on selected spectral bands, which leveragediscriminant information for mitosis classification on multispectral dataset. Sixth, we performa comprehensive study on region and patch based features for mitosis classification.Seven, we perform an extensive investigation of classifiers and inference of the best one formitosis classification. Eight, we propose an efficient and generic strategy to explore largeimages like WSI by combining computational geometry tools with a local signal measureof relevance in a dynamic sampling framework.The evaluation of these frameworks is done in MICO (COgnitive MIcroscopy, ANRTecSan project) platform prototyping initiative. We thus tested our proposed frameworks on MITOS international contest dataset initiated by this project. For the color framework,we manage to rank second during the contest. Furthermore, our multispectral frameworkoutperforms significantly the top methods presented during the contest. Finally, ourframeworks allow us reaching the same level of accuracy in mitosis detection on brightlightas multispectral datasets, a promising result on the way to clinical evaluation and routine.

Semantic segmentation as an approach to recognizing and localizing objects within an image is a major research area in computer vision. Now that convolutional neural networks are being increasingly used for such tasks, there have been many improve- ments in grand challenge results, and many new research opportunities in previously untennable areas. Using fully convolutional networks, we have developed a semantic segmentation pipeline for the identification of melanocytic tumor regions, epidermis, and dermis lay- ers in whole slide microscopy images of cutaneous melanoma or cutaneous metastatic melanoma. This pipeline includes processes for annotating and preparing a dataset from the output of a tissue slide scanner to the patch-based training and inference by an artificial neural network. We have curated a large dataset of 50 whole slide images containing cutaneous melanoma or cutaneous metastatic melanoma that are fully annotated at 40× ob- jective resolution by an expert pathologist. We will publish the source images of this dataset online. We also present two new FCN architectures that fuse multiple deconvolutional strides, combining coarse and fine predictions to improve accuracy over similar networks without multi-stride information. Our results show that the system performs better than our comparators. We include inference results on thousands of patches from four whole slide images, reassembling them into whole slide segmentation masks to demonstrate how our system generalizes on novel cases.

La formalisation des données cliniques est réalisée et adoptée dans plusieurs domaines de la santé comme la prévention des erreurs médicales, la standardisation, les guides de bonnes pratiques et de recommandations. Cependant, la communauté n'arrive pas encore à tirer pleinement profit de la valeur de ces données. Le problème majeur reste la difficulté à intégrer ces données et des services sémantiques associés au profit de la qualité de soins. Objectif L'objectif méthodologique de ce travail consiste à formaliser, traiter et intégrer les connaissances d'histopathologie et d'imagerie basées sur des protocoles standardisés, des référentiels et en utilisant les langages du web sémantique. L'objectif applicatif est de valoriser ces connaissances dans une plateforme pour faciliter l'exploration des lames virtuelles (LV), améliorer la collaboration entre pathologistes et fiabiliser les systèmes d'aide à la décision dans le cadre spécifique du diagnostic du cancer du sein. Il est important de préciser que notre but n'est pas de remplacer le clinicien, mais plutôt de l'accompagner et de faciliter ses lourdes tâches quotidiennes : le dernier mot reste aux pathologistes. Approche Nous avons adopté une approche transversale pour la représentation formelle des connaissances d'histopathologie et d'imagerie dans le processus de gradation du cancer. Cette formalisation s'appuie sur les technologies du web sémantique. / Semantic modelling of a histopathology image exploration and analysis tool. Recently, anatomic pathology (AP) has seen the introduction of several tools such as high-resolution histopathological slide scanners, efficient software viewers for large-scale histopathological images and virtual slide technologies. These initiatives created the conditions for a broader adoption of computer-aided diagnosis based on whole slide images (WSI) with the hope of a possible contribution to decreasing inter-observer variability. Beside this, automatic image analysis algorithms represent a very promising solution to support pathologist’s laborious tasks during the diagnosis process. Similarly, in order to reduce inter-observer variability between AP reports of malignant tumours, the College of American Pathologists edited 67 organ-specific Cancer Checklists and associated Protocols (CAP-CC&P). Each checklist includes a set of AP observations that are relevant in the context of a given organ-specific cancer and have to be reported by the pathologist. The associated protocol includes interpretation guidelines for most of the required observations. All these changes and initiatives bring up a number of scientific challenges such as the sustainable management of the available semantic resources associated to the diagnostic interpretation of AP images by both humans and computers. In this context, reference vocabularies and formalization of the associated knowledge are especially needed to annotate histopathology images with labels complying with semantic standards. In this research work, we present our contribution in this direction. We propose a sustainable way to bridge the content, features, performance and usability gaps between histopathology and WSI analysis.

Histopathology refers to inspecting and analysing tissue samples under a microscope to identify and examine signs of diseases. The manual investigation procedure of histology slides by pathologists is time-consuming and susceptible to misconceptions. Deep learning models have demonstrated outstanding performance in digital histopathology, providing doctors and clinicians with immediate and reliable decision-making assistance in their workflow. In this study, deep learning models, including vision transformers (ViT) and convolutional neural networks (CNN), were employed to compare their performance in patch-level classification task on feature annotations of glioblastoma multiforme in H\&E histology whole slide images (WSI). The dataset utilised in this study was obtained from the Ivy Glioblastoma Atlas Project (IvyGAP). The pre-processing steps included stain normalisation of the images, and patches of size 256x256 pixels were extracted from the WSIs. In addition, the per-subject split method was implemented to prevent data leakage between the training, validation and test sets. Three models were employed to perform the classification task on the IvyGAP data image, two scratch-trained models, a ViT and a CNN (variant of VGG16), and a pre-trained ViT. The models were assessed using various metrics such as accuracy, f1-score, confusion matrices, Matthews correlation coefficient (MCC), area under the curve (AUC) and receiver operating characteristic (ROC) curves. In addition, experiments were conducted to calibrate the models to reflect the ground truth of the task using the temperature scale technique, and their uncertainty was estimated through the Monte Carlo dropout approach. Lastly, the models were statistically compared using the Wilcoxon signed-rank test. Among the evaluated models, the scratch-trained ViT exhibited the best test accuracy of 67%, with an MCC of 0.45. The scratch-trained CNN obtained a test accuracy of 49% and an MCC of 0.15. However, the pre-trained ViT only achieved a test accuracy of 28% and an MCC of 0.034. The reliability diagrams and metrics indicated that the scratch-trained ViT demonstrated better calibration. After applying temperature scaling, only the scratch-trained CNN showed improved calibration. Therefore, the calibrated CNN was used for subsequent experiments. The scratch-trained ViT and calibrated CNN illustrated different uncertainty levels. The scratch-trained ViT had moderate uncertainty, while the calibrated CNN exhibited modest to high uncertainty across classes. The pre-trained ViT had an overall high uncertainty. Finally, the results of the statistical tests reported that the scratch-trained ViT model performed better among the three models at a significant level of approximately 0.0167 after applying the Bonferroni correction.  In conclusion, the scratch-trained ViT model achieved the highest test accuracy and better class discrimination. In contrast, the scratch-trained CNN and pre-trained ViT performed poorly and were comparable to random classifiers. The scratch-trained ViT demonstrated better calibration, while the calibrated CNN showed varying levels of uncertainty. The statistical tests demonstrated no statistical difference among the models.

Tesis por compendio / [ES] En los últimos años, el aprendizaje profundo (DL) se ha convertido en una de las principales áreas de la inteligencia artificial (IA), impulsado principalmente por el avance en la capacidad de procesamiento. Los algoritmos basados en DL han logrado resultados asombrosos en la comprensión y manipulación de diversos tipos de datos, incluyendo imágenes, señales de habla y texto. La revolución digital del sector sanitario ha permitido la generación de nuevas bases de datos, lo que ha facilitado la implementación de modelos de DL bajo el paradigma de aprendizaje supervisado. La incorporación de estos métodos promete mejorar y automatizar la detección y el diagnóstico de enfermedades, permitiendo pronosticar su evolución y facilitar la aplicación de intervenciones clínicas de manera más efectiva. Una de las principales limitaciones de la aplicación de algoritmos de DL supervisados es la necesidad de grandes bases de datos anotadas por expertos, lo que supone una barrera importante en el ámbito médico. Para superar este problema, se está abriendo un nuevo campo de desarrollo de estrategias de aprendizaje no supervisado o débilmente supervisado que utilizan los datos disponibles no anotados o débilmente anotados. Estos enfoques permiten aprovechar al máximo los datos existentes y superar las limitaciones de la dependencia de anotaciones precisas. Para poner de manifiesto que el aprendizaje débilmente supervisado puede ofrecer soluciones óptimas, esta tesis se ha enfocado en el desarrollado de diferentes paradigmas que permiten entrenar modelos con bases de datos débilmente anotadas o anotadas por médicos no expertos. En este sentido, se han utilizado dos modalidades de datos ampliamente empleadas en la literatura para estudiar diversos tipos de cáncer y enfermedades inflamatorias: datos ómicos e imágenes histológicas. En el estudio sobre datos ómicos, se han desarrollado métodos basados en deep clustering que permiten lidiar con las altas dimensiones inherentes a este tipo de datos, desarrollando un modelo predictivo sin la necesidad de anotaciones. Al comparar el método propuesto con otros métodos de clustering presentes en la literatura, se ha observado una mejora en los resultados obtenidos. En cuanto a los estudios con imagen histológica, en esta tesis se ha abordado la detección de diferentes enfermedades, incluyendo cáncer de piel (melanoma spitzoide y neoplasias de células fusocelulares) y colitis ulcerosa. En este contexto, se ha empleado el paradigma de multiple instance learning (MIL) como línea base en todos los marcos desarrollados para hacer frente al gran tamaño de las imágenes histológicas. Además, se han implementado diversas metodologías de aprendizaje, adaptadas a los problemas específicos que se abordan. Para la detección de melanoma spitzoide, se ha utilizado un enfoque de aprendizaje inductivo que requiere un menor volumen de anotaciones. Para abordar el diagnóstico de colitis ulcerosa, que implica la identificación de neutrófilos como biomarcadores, se ha utilizado un enfoque de aprendizaje restrictivo. Con este método, el coste de anotación se ha reducido significativamente al tiempo que se han conseguido mejoras sustanciales en los resultados obtenidos. Finalmente, considerando el limitado número de expertos en el campo de las neoplasias de células fusiformes, se ha diseñado y validado un novedoso protocolo de anotación para anotaciones no expertas. En este contexto, se han desarrollado modelos de aprendizaje profundo que trabajan con la incertidumbre asociada a dichas anotaciones. En conclusión, esta tesis ha desarrollado técnicas de vanguardia para abordar el reto de la necesidad de anotaciones precisas que requiere el sector médico. A partir de datos débilmente anotados o anotados por no expertos, se han propuesto novedosos paradigmas y metodologías basados en deep learning para abordar la detección y diagnóstico de enfermedades utilizando datos ómicos e imágenes histológicas. / [CA] En els últims anys, l'aprenentatge profund (DL) s'ha convertit en una de les principals àrees de la intel·ligència artificial (IA), impulsat principalment per l'avanç en la capacitat de processament. Els algorismes basats en DL han aconseguit resultats sorprenents en la comprensió i manipulació de diversos tipus de dades, incloent-hi imatges, senyals de parla i text. La revolució digital del sector sanitari ha permés la generació de noves bases de dades, la qual cosa ha facilitat la implementació de models de DL sota el paradigma d'aprenentatge supervisat. La incorporació d'aquests mètodes promet millorar i automatitzar la detecció i el diagnòstic de malalties, permetent pronosticar la seua evolució i facilitar l'aplicació d'intervencions clíniques de manera més efectiva. Una de les principals limitacions de l'aplicació d'algorismes de DL supervisats és la necessitat de grans bases de dades anotades per experts, la qual cosa suposa una barrera important en l'àmbit mèdic. Per a superar aquest problema, s'està obrint un nou camp de desenvolupament d'estratègies d'aprenentatge no supervisat o feblement supervisat que utilitzen les dades disponibles no anotades o feblement anotats. Aquests enfocaments permeten aprofitar al màxim les dades existents i superar les limitacions de la dependència d'anotacions precises. Per a posar de manifest que l'aprenentatge feblement supervisat pot oferir solucions òptimes, aquesta tesi s'ha enfocat en el desenvolupat de diferents paradigmes que permeten entrenar models amb bases de dades feblement anotades o anotades per metges no experts. En aquest sentit, s'han utilitzat dues modalitats de dades àmpliament emprades en la literatura per a estudiar diversos tipus de càncer i malalties inflamatòries: dades ómicos i imatges histològiques. En l'estudi sobre dades ómicos, s'han desenvolupat mètodes basats en deep clustering que permeten bregar amb les altes dimensions inherents a aquesta mena de dades, desenvolupant un model predictiu sense la necessitat d'anotacions. En comparar el mètode proposat amb altres mètodes de clustering presents en la literatura, s'ha observat una millora en els resultats obtinguts. Quant als estudis amb imatge histològica, en aquesta tesi s'ha abordat la detecció de diferents malalties, incloent-hi càncer de pell (melanoma spitzoide i neoplàsies de cèl·lules fusocelulares) i colitis ulcerosa. En aquest context, s'ha emprat el paradigma de multiple instance learning (MIL) com a línia base en tots els marcs desenvolupats per a fer front a la gran grandària de les imatges histològiques. A més, s'han implementat diverses metodologies d'aprenentatge, adaptades als problemes específics que s'aborden. Per a la detecció de melanoma spitzoide, s'ha utilitzat un enfocament d'aprenentatge inductiu que requereix un menor volum d'anotacions. Per a abordar el diagnòstic de colitis ulcerosa, que implica la identificació de neutròfils com biomarcadores, s'ha utilitzat un enfocament d'aprenentatge restrictiu. Amb aquest mètode, el cost d'anotació s'ha reduït significativament al mateix temps que s'han aconseguit millores substancials en els resultats obtinguts. Finalment, considerant el limitat nombre d'experts en el camp de les neoplàsies de cèl·lules fusiformes, s'ha dissenyat i validat un nou protocol d'anotació per a anotacions no expertes. En aquest context, s'han desenvolupat models d'aprenentatge profund que treballen amb la incertesa associada a aquestes anotacions. En conclusió, aquesta tesi ha desenvolupat tècniques d'avantguarda per a abordar el repte de la necessitat d'anotacions precises que requereix el sector mèdic. A partir de dades feblement anotades o anotats per no experts, s'han proposat nous paradigmes i metodologies basats en deep learning per a abordar la detecció i diagnòstic de malalties utilitzant dades *ómicos i imatges histològiques. Aquestes innovacions poden millorar l'eficàcia i l'automatització en la detecció precoç i el seguiment de malalties. / [EN] In recent years, deep learning (DL) has become one of the main areas of artificial intelligence (AI), driven mainly by the advancement in processing power. DL-based algorithms have achieved amazing results in understanding and manipulating various types of data, including images, speech signals and text. The digital revolution in the healthcare sector has enabled the generation of new databases, facilitating the implementation of DL models under the supervised learning paradigm. Incorporating these methods promises to improve and automate the detection and diagnosis of diseases, allowing the prediction of their evolution and facilitating the application of clinical interventions with higher efficacy. One of the main limitations in the application of supervised DL algorithms is the need for large databases annotated by experts, which is a major barrier in the medical field. To overcome this problem, a new field of developing unsupervised or weakly supervised learning strategies using the available unannotated or weakly annotated data is opening up. These approaches make the best use of existing data and overcome the limitations of reliance on precise annotations. To demonstrate that weakly supervised learning can offer optimal solutions, this thesis has focused on developing different paradigms that allow training models with weakly annotated or non-expert annotated databases. In this regard, two data modalities widely used in the literature to study various types of cancer and inflammatory diseases have been used: omics data and histological images. In the study on omics data, methods based on deep clustering have been developed to deal with the high dimensions inherent to this type of data, developing a predictive model without requiring annotations. In comparison, the results of the proposed method outperform other existing clustering methods. Regarding histological imaging studies, the detection of different diseases has been addressed in this thesis, including skin cancer (spitzoid melanoma and spindle cell neoplasms) and ulcerative colitis. In this context, the multiple instance learning (MIL) paradigm has been employed as the baseline in all developed frameworks to deal with the large size of histological images. Furthermore, diverse learning methodologies have been implemented, tailored to the specific problems being addressed. For the detection of spitzoid melanoma, an inductive learning approach has been used, which requires a smaller volume of annotations. To address the diagnosis of ulcerative colitis, which involves the identification of neutrophils as biomarkers, a constraint learning approach has been utilized. With this method, the annotation cost has been significantly reduced while achieving substantial improvements in the obtained results. Finally, considering the limited number of experts in the field of spindle cell neoplasms, a novel annotation protocol for non-experts has been designed and validated. In this context, deep learning models that work with the uncertainty associated with such annotations have been developed. In conclusion, this thesis has developed cutting-edge techniques to address the medical sector's challenge of precise data annotation. Using weakly annotated or non-expert annotated data, novel paradigms and methodologies based on deep learning have been proposed to tackle disease detection and diagnosis in omics data and histological images. These innovations can improve effectiveness and automation in early disease detection and monitoring. / The work of Rocío del Amor to carry out this research and to elaborate this dissertation has been supported by the Spanish Ministry of Universities under the FPU grant FPU20/05263. / Amor Del Amor, MRD. (2023). Deep Learning Strategies for Overcoming Diagnosis Challenges with Limited Annotations [Tesis doctoral]. Universitat Politècnica de València. https://doi.org/10.4995/Thesis/10251/200227 / Compendio

No description available.

Researchers within digital pathology are endeavouringto develop machine-learning tools to support dentists whenmaking a diagnosis. The purpose of this study was to investigatehow applying colour normalisation (CN) algorithms on an oral,histopathological dataset would impact both machine-learningmodels and ensembles of models when classifying cell types.The dataset was run through four different CN algorithms byusing a stain normalisation toolbox. The now five datasets (1 +4) were then fed separately into a pipeline to create machinelearningmodels, specifically convolutional neural networks withEfficientNet architecture. Two different ensembles were studied,one that used all the models and one that used the three modelswith the highest test accuracy. Each model gave a cell typeprediction of each cell. The ensembles super positioned theirmodels’ predictions of the same cell and used the results as theirown predictions.The models based on datasets created by two of the CNalgorithms had a weighted, average accuracy of ca. four percentagepoints higher than the model based on the unnormaliseddataset. Unexpectedly, the models based on the colour-normaliseddatasets had a larger standard deviation than the model basedon the unnormalised dataset. All the models were generally badat classifying two of the four cell types. Both the ensembleshad a weighted, average accuracy of ca. ten percentage pointshigher than the model based on the unnormalised dataset, aswell as a larger standard deviation. The increase in accuracyis significant and could move forward the timeline for whenmachine-leaning tools can be implemented into dentists’ andpathologists’ workflow. / Forskare inom digital patologi strävar efteratt utveckla maskininlärnings-verktyg som stödjer tandläkarenär de ställer diagnoser. Syftet med denna studie är att utreda hurtillämpning av färgnormaliserande algoritmer (CN algoritmer)på ett oralt, histopatologiskt dataset påverkar hur både maskininlärningsmodeller och ensembler av modeller klassificerarcelltyper.Datasetet kördes igenom fyra olika CN algoritmer med hjälpav en färgnormaliserings-verktygslåda. De nu fem dataseten(1 + 4) matades separat in i en ”pipeline” för att skapamaskininlärningsmodeller, specifikt djupa neurala nätverk medEfficientNet arkitektur. Två olika ensembler skapades, en somanvände alla modeller och en som endast använde de tre somhade högst noggrannhet på testsettet. Varje modell uppskattadecelltypen för varje cell. Ensemblerna superpositionerade derasmodellers uppskattningar för varje cell och använde resultatensom sina egna uppskattningar.Modellerna som tränats på två av de färgnormaliseradedataseten ökade i viktad, snitt-noggrannhet med fyra procentenheteri förhållande till modeller tränade på det ursprungligadatasetet. Förvånansvärt nog så ökade även standardavvikelsenhos modeller tränade på de färgnormaliserade dataseten. Allamodeller var generellt dåliga på att klassificera två av de fyracelltyperna. Ensemblen uppnådde en viktad snitt-noggrannhet på ca. tio procentenheter mer än modeller tränade på detursprungliga datasetet. Noggrannhetens signifikanta ökning kanleda till en tidigare implementering av maskininlärnings-verktygi tandläkares och patologers arbetsflöde. / Kandidatexjobb i elektroteknik 2021, KTH, Stockholm

With the increasing amount of information needed for diagnosis and therapeutic decision-making, and new trends such as “personalized medicine”, pathologists are expressing an increasing demand for automated tools that perform their most recurrent tasks in their daily practice, as well as an increase in the complexity of the analyses requested in their research activities. With current advances in histopathology, oncology, and biology, the current questions require the analysis of protein expression - evidenced using immunohistochemical (IHC) staining - within specific histological structures or tissue components, or the analysis of the co-expression of several proteins in a large number of tissue samples. In this Ph.D. thesis, we developed innovative solutions to make these analyses available for pathologists. To achieve this objective, we have used recent “machine learning” and, in particular, “deep learning” methodologies. We addressed different problems such as image normalization, to solve the important problem of inter-batch variability of IHC staining, and the automatic segmentation of histological structures, to compartmentalize protein expression quantification. Finally, we adapted image registration techniques to Tissue MicroArray (TMA) slide images to enable large-scale analyses of IHC staining colocalization. While imagenormalization will improve study reproducibility, the tools developed for automated segmentation will drastically reduce time and expert resources required for some studies as well as errors and imprecision due to the human factor. Finally, the work on image registration can provide answers to complex questions that require studying the potential interaction between several proteins on numerous histological samples. / Avec la quantité croissante d’informations nécessaires au diagnostic et à la prise de décision thérapeutique, et le développement de la “médecine personnalisée”, les pathologistes ont un besoin croissant d’outils automatisés pour exécuter leurs tâches les plus récurrentes. Ces outils se doivent également de réaliser des tâches de plus en plus complexes. En effet, avec les progrès récents en histopathologie, oncologie et biologie, les questions actuelles demandent, par exemple, l’analyse de l’expression de protéines révélées par marquages immunohistochimiques (IHC) au sein de structures ou compartiments histologiques spécifiques, ou encore l’analyse de la co-expression de plusieurs protéines dans un grand nombre d’échantillons. Dans cette thèse de doctorat, nous avons développé des solutions innovantes pour mettre ce type d’analyse à la disposition des pathologistes. Pour atteindre cet objectif, nous avons notamment fait appel à des méthodologies récentes de “machine learning” et, particulièrement, de “deep learning”. Nous avons ainsi abordé différentes questions telles que la normalisation d’images, pour résoudre l’important problème de la variabilité des marquages IHC, et la segmentation automatique de structures histologiques, pour permettre une quantification compartimentée de l’expression de protéines. Enfin, nous avons adapté des techniques dites de “recalage” aux images de lames de Tissue MicroArrays (TMA) pour permettre des analyses de colocalisation de marquages IHC à grande échelle. Alors que la normalisation des images améliore la reproductibilité des évaluations de marquages IHC, les outils développés pour la segmentation automatisée permettent de réduire significativement le temps et les ressources expertes nécessaires, ainsi que les erreurs et imprécisions dues au facteur humain. Enfin, les travaux sur le recalages d’images permettent d’apporter des éléments de réponse à des questions complexes qui nécessitent d’étudier l’interaction potentielle entre plusieurs protéines sur de nombreux échantillons histologiques. / Doctorat en Sciences de l'ingénieur et technologie / info:eu-repo/semantics/nonPublished

Proteins are the actual actors in the (normal or disrupted) physiological processes and immunohistochemistry (IHC) is a very efficient mean of visualizing and locating protein expression in tissue samples. By comparing pathologic and normal tissue, IHC is thus able to evidence protein expression alterations. This is the reason why IHC plays a grow- ing role to evidence tissue-based biomarkers in clinical pathology for diagnosing var- ious diseases and directing personalized therapy. Therefore, IHC biomarker evaluation significantly impacts the adequacy of the therapeutic choices for patients with serious pathologies, such as cancer. However, this evaluation may be time-consuming and dif- ficult to apply in practice due to the absence of precise positive cut-off values as well as staining (i.e. protein expression) heterogeneity intra- and inter-samples. Quantifying IHC staining patterns has thus become a crucial need in histopathology. For this task, automated image analysis has multiple advantages, such as avoiding the evidenced ef- fects of human subjectivity. The recent introduction of whole-slide scanners opened a wide range of possibilities for addressing challenging image analysis problems, includ- ing the identification of tissue-based biomarkers. Whole-slide scanners are devices that are able to image whole tissue slides at resolutions up to 0.1 micrometers per pixels, often referred to as virtual slides. In addition to quantification of IHC staining patterns, virtual slides are invaluable tools for the implementation of digital pathology work- flows. The present work aims to make several contributions towards this current digital shift in pathology. Our first contribution was to propose an automated virtual slide sharpness assessment tool. Although modern whole-slide scanner devices resolve most image standardization problems, focusing errors are still likely to be observed, requiring a sharpness assessment procedure. Our proposed tool will ensure that images provided to subsequent pathologist examination and image analysis are correctly focused. Virtual slides also enable the characterization of biomarker expression heterogeneity. Our sec- ond contribution was to propose a method to characterize the distribution of densely stained regions in the case of nuclear IHC biomarkers, with a focus on the identification of highly proliferative tumor regions by analyzing Ki67-stained tissue slides. Finally, as a third contribution, we propose an efficient mean to register virtual slides in order to characterize biomarker colocalization on adjacent tissue slides. This latter contribution opens new prospects for the analysis of more complex questions at the tissue level and for finely characterizing disease processes and/or treatment responses./Les protéines sont les véritables acteurs des processus physiologiques (normaux ou per- turbés) et l’immunohistochimie (IHC) est un moyen efficace pour visualiser et localiser leur expression au sein d’échantillons histologiques. En comparant des échantillons de tissus pathologiques et normaux, l’IHC permet de révéler des altérations dans des pro- fils d’expression protéique. C’est pourquoi l’IHC joue un rôle de plus en plus important pour mettre en évidence des biomarqueurs histologiques intervenant dans le diagnos- tic de diverses pathologies et dans le choix de thérapies personnalisées. L’évaluation de l’expression de biomarqueurs révélés par IHC a donc des répercussions importantes sur l’adéquation des choix thérapeutiques pour les patients souffrant de pathologies graves, comme le cancer. Cependant, cette évaluation peut être chronophage et difficile à appliquer en pratique, d’une part, à cause de l’hétérogénéité de l’expression protéique intra- et inter-échantillon, d’autre part, du fait de l’absence de critères de positivité bien définis. Il est donc devenu crucial de quantifier les profils d’expression de marquages IHC en histopathologie. A cette fin, l’analyse d’image automatisée possède de multiples avantages, comme celui d’éviter les effets de la subjectivité humaine, déjà démontrés par ailleurs. L’apparition récente des numériseurs de lames histologiques complètes, ou scanners de lames, a permis l’émergence d’un large éventail de possibilités pour traiter des problèmes d’analyse d’image difficiles menant à l’identification de biomar- queurs histologiques. Les scanners de lames sont des dispositifs capables de numériser des lames histologiques à une résolution pouvant atteindre 0,1 micromètre par pixel, expliquant la dénomination de "lames virtuelles" des images ainsi acquises. En plus de permettre la quantification des marquages IHC, les lames virtuelles sont des outils indis- pensables pour la mise en place d’un flux de travail numérique en pathologie. Le travail présenté ici vise à fournir plusieurs contributions au récent changement de cap vers une numérisation de la discipline médicale qu’est l’anatomie pathologique. Notre première contribution consiste en un outil permettant d’évaluer automatiquement la netteté des lames virtuelles. En effet, bien que les scanners de lames résolvent la plupart des pro- blèmes liés à la standardisation de l’acquisition, les erreurs de focus restent fréquentes, ce qui nécessite la mise en place d’une procédure de vérification de la netteté. L’outil que nous proposons assurera la netteté des images fournies à l’examen du pathologiste et à l’analyse d’image. Les lames virtuelles permettent aussi de caractériser l’hétérogénéité de l’expression de biomarqueurs. Ainsi, la deuxième contribution de ce travail repose sur une méthode permettant de caractériser la distribution de régions densément marquées par des biomarqueurs IHC nucléaires. Pour ce travail, nous nous sommes concentrés sur l’identification de régions tumorales présentant une forte activité proliférative en analysant des lames virtuelles révélant l’expression de la protéine Ki67. Finalement, la troisième contribution de ce travail fut de proposer un moyen efficace de recaler des lames virtuelles dans le but de caractériser la colocalisation de biomarqueurs IHC révé- lés sur des coupes de tissu adjacentes. Cette dernière contribution ouvre de nouvelles perspectives pour l’analyse de questions complexes au niveau histologique ainsi que la caractérisation fine de processus pathologiques et de réponses thérapeutiques. / Doctorat en Sciences de l'ingénieur / info:eu-repo/semantics/nonPublished

La pathologie numérique s’est développée ces dernières années grâce à l’avancée récente des algorithmes d’analyse d’images et de la puissance de calcul. Notamment, elle se base de plus en plus sur les images histologiques. Ce format de données a la particularité de révéler les objets biologiques recherchés par les experts en utilisant des marqueurs spécifiques tout en conservant la plus intacte possible l’architecture du tissu. De nombreuses méthodes d’aide au diagnostic à partir de ces images se sont récemment développées afin de guider les pathologistes avec des mesures quantitatives dans l’établissement d’un diagnostic. Les travaux présentés dans cette thèse visent à adresser les défis liés à l’analyse d’images histologiques, et à développer un modèle d’aide au diagnostic se basant principalement sur les relations spatiales, une information que les méthodes existantes n’exploitent que rarement. Une technique d’analyse de la texture à plusieurs échelles est tout d’abord proposée afin de détecter la présence de tissu malades dans les images. Un descripteur d’objets, baptisé Force Histogram Decomposition (FHD), est ensuite introduit dans le but d’extraire les formes et l’organisation spatiale des régions définissant un objet. Finalement, les images histologiques sont décrites par les FHD mesurées à partir de leurs différents types de tissus et des objets biologiques marqués qu’ils contiennent. Les expérimentations intermédiaires ont montré que les FHD parviennent à correctement reconnaitre des objets sur fonds uniformes y compris dans les cas où les relations spatiales ne contiennent à priori pas d’informations pertinentes. De même, la méthode d’analyse de la texture s’avère satisfaisante dans deux types d’applications médicales différents, les images histologiques et celles de fond d’œil, et ses performances sont mises en évidence au travers d’une comparaison avec les méthodes similaires classiquement utilisées pour l’aide au diagnostic. Enfin, la méthode dans son ensemble a été appliquée à l’aide au diagnostic pour établir la sévérité d’un cancer via deux ensembles d’images histologiques, un de foies métastasés de souris dans le contexte du projet ANR SPIRIT, et l’autre de seins humains dans le cadre du challenge CPR 2014 : Nuclear Atypia. L’analyse des relations spatiales et des formes à deux échelles parvient à correctement reconnaitre les grades du cancer métastasé dans 87, 0 % des cas et fourni des indications quant au degré d’atypie nucléaire. Ce qui prouve de fait l’efficacité de la méthode et l’intérêt d’encoder l’organisation spatiale dans ce type d’images particulier. / During the last decade, digital pathology has been improved thanks to the advance of image analysis algorithms and calculus power. Particularly, it is more and more based on histology images. This modality of images presents the advantage of showing only the biological objects targeted by the pathologists using specific stains while preserving as unharmed as possible the tissue structure. Numerous computer-aided diagnosis methods using these images have been developed this past few years in order to assist the medical experts with quantitative measurements. The studies presented in this thesis aim at adressing the challenges related to histology image analysis, as well as at developing an assisted diagnosis model mainly based on spatial relations, an information that currently used methods rarely use. A multiscale texture analysis is first proposed and applied to detect the presence of diseased tissue. A descriptor named Force Histogram Decomposition (FHD) is then introduced in order to extract the shapes and spatial organisation of regions within an object. Finally, histology images are described by the FHD measured on their different types of tissue and also on the stained biological objects inside every types of tissue. Preliminary studies showed that the FHD are able to accurately recognise objects on uniform backgrounds, including when spatial relations are supposed to hold no relevant information. Besides, the texture analysis method proved to be satisfactory in two different medical applications, namely histology images and fundus photographies. The performance of these methods are highlighted by a comparison with the usual approaches in their respectives fields. Finally, the complete method has been applied to assess the severity of cancers on two sets of histology images. The first one is given as part of the ANR project SPIRIT and presents metastatic mice livers. The other one comes from the challenge ICPR 2014 : Nuclear Atypia and contains human breast tissues. The analysis of spatial relations and shapes at two different scales achieves a correct recognition of metastatic cancer grades of 87.0 % and gives insight about the nuclear atypia grade. This proves the efficiency of the method as well as the relevance of measuring the spatial organisation in this particular type of images.