Dissertationen zum Thema „Histopathologie digitale“

Un défi majeur dans l'application de l'apprentissage profond à l'histopathologie réside dans la variation des colorations, à la fois inter et intra-coloration. Les modèles d'apprentissage profond entraînés sur une seule coloration (ou domaine) échouent souvent sur d'autres, même pour la même tâche (par exemple, la segmentation des glomérules rénaux). L'annotation de chaque coloration est coûteuse et chronophage, ce qui pousse les chercheurs à explorer des méthodes de transfert de coloration basées sur l'adaptation de domaine. Celles-ci visent à réaliser une segmentation multi-coloration en utilisant des annotations d'une seule coloration, mais sont limitées par l'introduction d'un décalage de domaine, réduisant ainsi les performances. La détection et la quantification de ce décalage sont essentielles. Cette thèse se concentre sur des méthodes non supervisées pour développer une métrique de détection du décalage et propose une approche de transfert de coloration pour le minimiser. Bien que ces algorithmes réduisent le besoin d'annotations, ils peuvent être limités pour certains tissus. Cette thèse propose donc une amélioration via l'auto-supervision
A key challenge in applying deep learning to histopathology is the variation in stainings, both inter and intra-stain. Deep learning models trained on one stain (or domain) often fail on others, even for the same task (e.g., kidney glomeruli segmentation). Labelling each stain is expensive and time-consuming, prompting researchers to explore domain adaptation based stain-transfer methods. These aim to perform multi-stain segmentation using labels from only one stain but are limited by the introduction of domain shift, reducing performance. Detecting and quantifying this domain shift is important. This thesis focuses on unsupervised methods to develop a metric for detecting domain shift and proposes a novel stain-transfer approach to minimise it. While multi-stain algorithms reduce the need for labels in target stains, they may struggle with tissue types lacking source-stain labels. To address this, the thesis focuses to improve multi-stain segmentation with less reliance on labelled data using self-supervision. While this thesis focused on kidney glomeruli segmentation, the proposed methods are designed to be applicable to other histopathology tasks and domains, including medical imaging and computer vision

L’angiogenèse est le processus par lequel de nouveaux vaisseaux sanguins se forment à partir du réseaux préexistant. Au cours de l’angiogenèse tumorale, les cellules tumorales sécrètent des facteurs de croissance qui activent la prolifération et la migration des cellules et stimulent la surproduction du facteur de croissance endothélial vasculaire (VEGF). Le rôle fondamental de l’approvisionnement vasculaire dans la croissance tumorale et le developement des thérapies anticancéreuses rend l’évaluation de l’angiogenèse tumorale, cruciale dans l’évaluation de l’effet des thérapies anti-angiogéniques, en tant que thérapie anticancéreuse prometteuse. Dans cette étude, nous établissons un panel quantitatif et qualitatif pour évaluer les structures des vaisseaux sanguins de la tumeur sur des images de fluorescence non invasives et des images histopathologique sur toute la surface tumorale afin d’identifier les caractéristiques architecturales et les mesures quantitatives souvent associées à la réponse thérapeutique ou prédictive de celle-ci. Nous développons un pipeline formé de Markov Random Field (MFR) et Watershed pour segmenter les vaisseaux sanguins et les composants du micro-environnement tumoral afin d’évaluer quantitativement l’effet du médicament anti-angiogénique Pazopanib sur le système vasculaire tumoral et l’interaction avec le micro-environnement de la tumeur. Le pazopanib, agent anti-angiogénèse, a montré un effet direct sur le système vasculaire du réseau tumoral via les cellules endothéliales. Nos résultats montrent une relation spécifique entre la néovascularisation apoptotique et la densité de noyau dans une tumeur murine traitée par Pazopanib. Une évaluation qualitative des vaisseaux sanguins de la tumeur est réalisée dans la suite de l’étude. Nous avons développé un modèle de réseau de neurone discriminant-générateur basé sur un modele d’apprentissage : réseau de neurones convolutionnels (CNN) et un modèle de connaissance basé sur des règles Marked Point Process (MPP) permettant de segmenter les vaisseaux sanguins sur des images très hétérogènes à l’aide de très peu de données annotées. Nous détaillons l’intuition et la conception du modèle discriminatif-génératif, sa similarité avec les Réseaux antagonistes génératifs (GAN) et nous évaluons ses performances sur des données histopathologiques et synthétiques. Les limites et les perspectives de la méthode sont présentées à la fin de notre étude
Angiogenesis is the process through which new blood vessels are formed from pre-existing ones. During angiogenesis, tumour cells secrete growth factors that activate the proliferation and migration of endothelial cells and stimulate over production of the vascular endothelial growth factor (VEGF). The fundamental role of vascular supply in tumour growth and anti-cancer therapies makes the evaluation of angiogenesis crucial in assessing the effect of anti-angiogenic therapies as a promising anti-cancer therapy. In this study, we establish a quantitative and qualitative panel to evaluate tumour blood vessels structures on non-invasive fluorescence images and histopathological slide across the full tumour to identify architectural features and quantitative measurements that are often associated with prediction of therapeutic response. We develop a Markov Random Field (MFRs) and Watershed framework to segment blood vessel structures and tumour micro-enviroment components to assess quantitatively the effect of the anti-angiogenic drug Pazopanib on the tumour vasculature and the tumour micro-enviroment interaction. The anti-angiogenesis agent Pazopanib was showing a direct effect on tumour network vasculature via the endothelial cells crossing the whole tumour. Our results show a specific relationship between apoptotic neovascularization and nucleus density in murine tumor treated by Pazopanib. Then, qualitative evaluation of tumour blood vessels structures is performed in whole slide images, known to be very heterogeneous. We develop a discriminative-generative neural network model based on both learning driven model convolutional neural network (CNN), and rule-based knowledge model Marked Point Process (MPP) to segment blood vessels in very heterogeneous images using very few annotated data comparing to the state of the art. We detail the intuition and the design behind the discriminative-generative model, and we analyze its similarity with Generative Adversarial Network (GAN). Finally, we evaluate the performance of the proposed model on histopathology slide and synthetic data. The limits of this promising framework as its perspectives are shown

Histological analysis of tissue biopsies by an expert pathologist is considered gold standard for diagnosing many cancers, including breast cancer. Nottingham grading system, which is the most widely used criteria for histological grading of breast tissues, consists of three components: mitotic count, nuclear atypia and tubular formation. In routine histological analysis, pathologists perform grading of breast cancer tissues by manually examining each tissue specimen against the three components, which is a laborious and subjective process and thus can suffer from low inter-observer agreement. With the advent of digital whole-slide scanning platforms, automatic image analysis algorithms can be used as a partial solution for these issues. The main goal of this dissertation is to develop frameworks that can aid towards building an automated or semi-automated breast cancer grading system. We present novel frameworks for detection of mitotic cells and nuclear atypia scoring in breast cancer histopathology images. Both of these frameworks can play a fundamental role in developing a computer-assisted breast cancer grading system. Moreover, the proposed image analysis frameworks can be adapted to grading and analysis of cancers of several other tissues such as lung and ovarian cancers. In order to deal with one of the fundamental problems in histological image analysis applications, we first present a stain normalisation algorithm that minimises the staining inconsistency in histological images. The algorithm utilises a novel image-specific colour descriptor which summarises the colour contents of a histological image. Stain normalisation algorithm is used in the remainder of the thesis as a preprocessing step. We present a mitotic cell detection framework mimicking a pathologist’s approach, whereby we first perform tumour segmentation to restrict our search for mitotic cells to tumour regions only, followed by candidate detection and evaluation in a statistical machine learning framework. We also employ a discriminative dictionary learning paradigm to learn the visual appearance of mitotic cells, that models colour, texture, and shape in a composite manner. Finally, we present a nuclear atypia scoring framework based on a novel image descriptor which summarises the texture heterogeneity, inherent in histological images in a compact manner. Classification is performed using a geodesic k-nearest neighbour classifier which explicitly exploits the structure of Riemannian manifold of the descriptor and achieves significant performance boost as compared to Euclidean counterpart.

This thesis focuses on developing new automatic techniques addressing three typical problems in digital histopathology image analysis, histochemical stain separation at pixel-level, cell classifications at region level, and histochemical score assessment at image level, with the aim of providing useful tools to help histopathologists in their decision making. First, we study a pixel-level problem, separating positive chemical stains. To realise the full potential of digital pathology, accurate and robust computer techniques for automatically detecting biomarkers play an important role. Traditional methods transform the colour histopathology images into a gray scale image and apply a single threshold to separate positively stained tissues from the background. In this thesis, we show that the colour distribution of the positive immunohistochemical stains varies with the level of luminance and that a single threshold will be impossible to separate positively stained tissues from other tissues, regardless how the colour pixels are transformed. Based on this observation, two novel luminance adaptive biomarker detection methods are proposed. The first, termed Luminance Adaptive Multi-Thresholding (LAMT) first separate the pixels according to their luminance levels and for each luminance level a separate threshold is found for detecting the positive stains. The second, termed Luminance Adaptive Random Forest (LARF) applies one of the most powerful machine learning models, random forest, as a base classifier to build an ensemble classifier for biomarker detection. Second, we study a cell-level problem, the cell classification task in pathology images. Two different classification models are proposed. The first model for HEp-2 cell pattern classification comes with a novel object-graph based feature, which decompose the binary image into primitive objects and represent them with a set of morphological feature. Work on cell classification is further extended using deep learning model termed Deep Autoencoding-Classification Network (DACN). The DACN model consists of an autoencoder and a conventional classification convolutional neural network (CNN) with the two sharing the same encoding pipeline. The DACN model is jointly optimized for the classification error and the image reconstruction error based on a multi-task learning procedure. We will present experiment results to show that the proposed DACN outperforms all known state-of-the-art on two public indirect immunofluorescence stained HEp-2 cell datasets and H\&E stained colorectal adenocarcinomas cell dataset. Third, we study an image-level problem, assessing the histochemical score of a histopathology image. To determine the molecular class of the tumour, pathologists will have to manually mark the nuclei activity biomarkers by assigning a histochemical score (H-Score) to each TMA core with a semi-quantitative assessment method. Manually marking positively stained nuclei is a time consuming, imprecise and subjective process which will lead to inter-observer and intra-observer discrepancies. In this thesis, we present an end-to-end deep learning system which directly predicts the H-Score automatically. Our system imitates the pathologists' decision process and uses one fully convolutional network (FCN) to extract all nuclei region, a second FCN to extract tumour nuclei region, and a multi-column convolutional neural network which takes the outputs of the first two FCNs and the stain intensity description image as input and acts as the decision making mechanism to directly output the H-Score of the input TMA image. To the best of our knowledge, this is the first end-to-end system that takes a TMA image as input and directly outputs a clinical score. We will present experimental results which demonstrate that the H-Scores predicted by our model have very high and statistically significant correlation with experienced pathologists' scores and that the H-Score discrepancy between our algorithm and the pathologists is on par with the inter-subject discrepancy between the pathologists.

In 2012, more than 1.6 million new cases of breast cancer were diagnosed and about half a million women died of breast cancer. The incidence has increased in the developing world. The mortality, however, has decreased. This is thought to partly be the result of advances in diagnosis and treatment. Studying tissue samples from biopsies through a microscope is an important part of diagnosing breast cancer. Recent techniques include camera-equipped microscopes and whole slide scanning systems that allow for digital high-throughput scanning of tissue samples. The introduction of digital pathology has simplified parts of the analysis, but manual interpretation of tissue slides is still labor intensive and costly, and involves the risk for human errors and inconsistency. Digital image analysis has been proposed as an alternative approach that can assist the pathologist in making an accurate diagnosis by providing additional automatic, fast and reproducible analyses. This thesis addresses the automation of conventional analyses of tissue, stained for biomarkers specific for the diagnosis of breast cancer, with the purpose of complementing the role of the pathologist. In order to quantify biomarker expression, extraction and classification of sub-cellular structures are needed. This thesis presents a method that allows for robust and fast segmentation of cell nuclei meeting the need for methods that are accurate despite large biological variations and variations in staining. The method is inspired by sparse coding and is based on dictionaries of local image patches. It is implemented in a tool for quantifying biomarker expression of various sub-cellular structures in whole slide images. Also presented are two methods for classifying the sub-cellular localization of staining patterns, in an attempt to automate the validation of antibody specificity, an important task within the process of antibody generation.  In addition, this thesis explores methods for evaluation of multimodal data. Algorithms for registering consecutive tissue sections stained for different biomarkers are evaluated, both in terms of registration accuracy and deformation of local structures. A novel region-growing segmentation method for multimodal data is also presented. In conclusion, this thesis presents computerized image analysis methods and tools of potential value for digital pathology applications.

Oral cavity cancer (OCC) is a type of cancer of the lip, tongue, salivary glands and other sites in the mouth (buccal or oral cavity) and is the sixth leading cause of cancer worldwide. Patients with OCC are treated based on a staging system: low-stage patients typically receive less aggressive therapy compared to high-stage patients. Unfortunately, low-stage patients are sometimes at risk for locoregional recurrence. Recently, a semi-quantitative risk scoring system has been developed to assess the locoregional recurrence risk for low-stage patients. This risk scoring system is based on tissue characteristics determined on 2D histopathology images under a microscope. This modality limits the appreciation of the 3D architecture of the tumor and its associated morphological features. This thesis aims to visualize 3D models of the tumor-host interface reconstructed from serially-sectioned histopathology slides and quantify their clinically validated morphological features to predict locoregional recurrence after treatment. The 3D models are developed and quantified for 6 patient cases using readily available tools. This pilot study provides a framework for an automated diagnostic technique for 3D visualization and morphological analysis of tumor biology which is traditionally done using 2D analysis.

Histopathology refers to the visual inspection of tissue under the microscope and it is the core part of diagnosis. The process of manual inspection of histopathology slides is very time-consuming for pathologists and error-prone. Furthermore, diagnosis can sometimes differ among specialists. In recent years, convolutional neural networks (CNNs) have demonstrated remarkable performances in the classification of digital histopathology images. However, due to their high resolution, whole-slide images are of immense size, often gigapixels, making it infeasible to train CNNs directly on them. For that, patch-level classification is used instead. In this study, a deep learning approach for patch-level classification of glioblastoma (i.e. brain cancer) tissue is proposed. Four different state-of-the-art models were evaluated (MobileNetV2, ResNet50, ResNet152V2, and VGG16), with MobileNetV2 being the best among them, achieving 80% test accuracy. The study also proposes a scratch-trained CNN architecture, inspired by the popular VGG16 model, which achieved 81% accuracy. Both models scored 87% test accuracy when trained with data augmentation. All models were trained and tested on randomly sampled patches from the Ivy GAP dataset, which consisted of 724 H&E images in total. Finally, as patch-level predictions cannot be used explicitly by pathologists, prediction results from two slides were presented in the form of whole-slide images. Post-processing was also performed on those two predicted WSIs in order to make use of the spatial correlations among the patches and increase the classification accuracy. The models were statistically compared using the Wilcoxon signed-rank test.

Le 21ème siècle a vu l'essor de la pathologie numérique. De fait, les challenges de l'analyse des données histopathologiques ont contribué à un effort mondial dans la lutte globale contre le cancer. Parallèlement, le succès récent de la décision par automate, plus particulièrement l'apprentissage profond, a révolutionné la recherche dans le domaine de la vision par ordinateur. Dans cette thèse, nous avons étudié la prédiction de la réponse au traitement chez des patients atteints d'un cancer du sein triple négatif avec deux approches différentes aux performances similaires. La première approche, basée sur le récent succès de la vision par ordinateur, extrait des caractéristiques afin d'en effectuer la classification finale. La deuxième approche contraint le flux d'information à passer par la segmentation de noyaux. En particulier, elle permet d'incorporer des informations de haute résolution à une vue globale basse résolution. Bien que cette approche soit plus attrayante, puisqu'elle repose sur l'analyse et la quantification d'un élément biologique précis, la segmentation de noyaux est une tâche fastidieuse. Nous proposons une nouvelle approche de segmentation par apprentissage profond, qui est particulièrement adaptée à la séparation de cellules en contact
The rise of digital pathology and with it the challenges of histopathology analysis have been the focus of a worldwide effort in the overall fight against cancer. In parallel, the recent success of automated decision-making, machine learning, and specifically deep learning, have revolutionised the basis of research as we know today. In this thesis, we tackle the prediction of treatment response in triple-negative breast cancer patients with two different approaches that reach similar outcomes. The first line of approach, based on the recent success of computer vision, extracts learned features from the data in order to perform classification. The second line of approach forces the information flow to pass through nuclei segmentation. In particular, it allows the incorporation of high-resolution information on to a lower resolution overview. Yet while this approach is more appealing as it is based on the analysis and quantification of a precise biological element, nuclei segmentation is troublesome. While solving the task of nuclei segmentation with deep learning, we propose a new formulation for nuclei segmentation which excels at separating touching objects

Renal clear cell carcinoma affects the kidneys by abnormal cell division which spreads to other organs through the bloodstream and lymphatic system. The number of renal cancer cases grows whilst rapid and accurate diagnoses are required for early intervention. Biopsies are critical for cancer diagnosis. Pathologists look beyond manual evaluation to include computer-based analysis to develop accurate cancer diagnostics. Pathologists render diagnostic reports to assist with treatment whilst expert analysis is time consuming and restricts early diagnosis. The process of manual expert pathology reporting is prohibitive and poor and repetitive concentration can lead to misdiagnosis. The probability of observational error increases along with the increased workload of the average pathologist and the demand for histopathology image analysis. Advances in computational and computer-assisted applications can provide accurate and timely analysis of histopathology images. Manual annotation now looks to machine learning algorithms. The nuclei segmentation technique is one possible approach. It uses deep learning-based nuclei segmentation approaches to train networks. This assists expert pathology which is expensive and time-consuming. Overlapping nuclei segmentation is a challenging issue for automated histopathology image analysis. Deep observation is required in digital images to identify the overlapping nuclei and variations in segmentation errors can mislead expert pathologists. The aim of this research study was to perform a literature review of existing nuclei segmentation techniques including overlapping splitting algorithms; identify the limitations and knowledge gaps; and propose computerised deep learning based individual nuclei segmentation and analysis of histopathology images. A mixed method research study was performed with sequential research experiments in data collection; image pre-processing; synthetic image generation; segmentation of nuclei regions; overlapping nuclei; and the validation of a proposed framework. A series of experiments were executed to find the most viable approach. An improved approach was designed for synthetic image generation using a cycle-consistent GAN network. The network created synthetic backgrounds and allowed for a CNN filtering method to separate the initial synthetic backgrounds. Nuclei shapes were collected to create transformed shapes. These transformed shapes were placed on the refined synthetic backgrounds to generate complete synthetic images. The similarity of original and synthetic images established and viable, valid pathway. A nuclei mask of synthetic images was collected to train a modified U-net segmentation network for better segmentation accuracy. These synthetic images performed better than original images. Accurately delineating the individual nucleus boundary helped to generatean automated system divide the nuclei clumps into individual nuclei in histopathology images. Using the nuclei ground truth of original images, it was possible to validate an application that informed manual expert pathology and to process multiple images and minimise histopathology image analysis. The novelty of this research is the creation of an automated deep learning based individual nuclei segmentation system for renal histopathology images. The synthetic images and corresponding nuclei masks were trained with a modified U-net nuclei segmentation network. The trained network provides better nuclei segmentation performance in original images. The research developed a robust application which allows for the analysis of multiple histopathology images.

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

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

Brain and nervous system tumors were responsible for around 250,000 deaths in 2020 worldwide. Correctly identifying different tumors is very important, because treatment options largely depend on the diagnosis. This is an expert task, but recently machine learning, and especially deep learning models have shown huge potential in tumor classification problems, and can provide fast and reliable support for pathologists in the decision making process. This thesis investigates classification of two brain tumors, glioblastoma multiforme and lower grade glioma in high-resolution H&E-stained histology images using deep learning. The dataset is publicly available from TCGA, and 220 whole slide images were used in this study. Ground truth labels were only available on whole slide level, but due to their large size, they could not be processed by convolutional neural networks. Therefore, patches were extracted from the whole slide images in two sizes and fed into separate networks for training. Preprocessing steps ensured that irrelevant information about the background was excluded, and that the images were stain normalized. The patch-level predictions were then combined to slide level, and the classification performance was measured on a test set. Experiments were conducted about the usefulness of pre-trained CNN models and data augmentation techniques, and the best method was selected after statistical comparisons. Following the patch-level training, five slide aggregation approaches were studied, and compared to build a whole slide classifier model. Best performance was achieved when using small patches (336 x 336 pixels), pre-trained CNN model without frozen layers, and mirroring data augmentation. The majority voting slide aggregation method resulted in the best whole slide classifier with 91.7% test accuracy and 100% sensitivity. In many comparisons, however, statistical significance could not be shown because of the relatively small size of the test set.

Thesis (MMed)--Stellenbosch University, 2015.
ENGLISH ABSTRACT: Background: The LODOX Statscan is a whole-body digital X-ray scanning device which was adapted for medical usage. The LODOX has an established role in the field of Forensic Pathology where it shows high sensitivity and specificity for the detection of skeletal pathology and foreign bodies. The role of the scanner in the detection of soft tissue pathology in the lungs of adults has not been reported and this study aims to review the radio-pathological correlation and the applicability of LODOX as a viable screening tool in the detection of lung pathology in post mortem cases. Methods: We prospectively reviewed cases which were referred for medico-legal autopsy between November 2012 and March 2013 to the Tygerberg Forensic Pathology Service mortuary, Cape Town, South Africa. All cases meeting the prescribed inclusion criteria underwent LODOX scanning as well as macroscopic and microscopic evaluation of the lungs as permitted by the Inquests Act 58 of 1959. The macroscopic and microscopic variables were considered the “gold standard” when compared with the results of the LODOX. The sensitivity, specificity, positive and negative predictive values were assessed. Results: One hundred and fifty nine cases (159) were included in the study. The most common radiographic patterns reported were the presence of ground glass opacities and consolidation. Overall, low to moderate sensitivity of these LODOX patterns in the prediction of pneumonic microscopic pathology (oedema, acute and chronic inflammation and features of diffuse alveolar damage) was noted. These values were lower than that reported for pneumonia using conventional X-rays. Additionally, these LODOX patterns have a high probability of representing oedema or autolytic/decomposition change. Pneumothorax was the most common pleural pathology detected on LODOX, but autopsy correlation could not be performed. Poor to no correlation was noted with the variables of cavity, malignant tumour, and bronchiectasis, but the prevalence of these conditions in our cohort was low. In general, LODOX predictions were better at excluding pathology which was not present rather than confirming pathology which was present. Conclusions: The LODOX offers excellent evidentiary value in the demonstration of a pneumothorax but currently has limited value as a “stand alone” test in the field of Forensic Pathology. However the continued use of the LODOX as an adjunct examination, as well as prospective study of its applicability, is advised.
AFRIKAANSE OPSOMMING: Agtergrond: Die LODOX Statscan is ‘n heel-liggaam digitale X-straal skandeer apparaat wat aangepas is vir mediese gebruik. Die LODOX het ‘n gevestigde rol in Geregtelike Patologie, waar dit ‘n hoë sensitiwiteit en spesifisiteit het in die opsporing van skeletale patologie en vreemde voorwerpe. Die rol van die skandeerder in die opspoor van sagte weefsel patologie in die longe van volwassenes is nog nie gerapporteer nie, en hierdie studie ondersoek die radio-patologiese korrelasie en toepaslikheid van LODOX as ‘n doeltreffende siftingsmeganisme om long patologie op te spoor in post-mortale gevalle. Metode: Gevalle wat verwys is na die Tygerberg Geregtelike Patologie Diens lykshuis in Kaapstad, Suid-Afrika vir medies-geregtelike outopsies tussen November 2012 en Maart 2013, is prospektief geëvalueer. Alle gevalle wat die voorgeskrewe insluitingskriteria nagekom het, het LODOX skandering asook makroskopiese en mikroskopiese ondersoek van die longe ondergaan, soos toegelaat deur die Wet op Geregtelike Doodsondersoeke Nr 58 van 1959. Die makroskopiese en mikroskopiese veranderlikes is beskou as die “goud standaard” in vergelyking met die resultate van die LODOX. Die sensitiwiteit, spesifisiteit, positiewe en negatiewe voorspellingswaardes is beoordeel. Resultate: Eenhonderd-nege-en-vyftig gevalle (159) is ingesluit in die studie. Die algemeenste radiografiese pattroon wat gerapporteer is, was die teenwoordigheid van gemaalde glas opasiteit en konsolidasie. In geheel is lae to matige sensitiwiteit van hierdie LODOX beelde waargeneem in die voorspelling van pneumoniese mikroskopiese patologie (edeem, akute en chroniese ontsteking, en eienskappe van diffuse alveolêre skade). Hierdie waardes was laer as die wat gerapporteer is vir pneumonie met konvensionele X-strale. Verder het hierdie LODOX beelde ‘n hoë waarskynlikheid om edeem en/of outolise/ontbinding uit te beeld. Pneumotoraks was die algemeenste pleurale patologie wat waargeneem is met die LODOX, maar outopsie korrelasie kon nie gedoen word nie. Swak tot geen korrelasie is gemerk vir die veranderlikes kaviteit, maligne tumor en brongi-ektase, maar die prevalensie van hierdie toestande in ons kohort was laag. Oor die algemeen was LODOX voorspellings beter om patologie wat nie teenwoordig is nie, uit te skakel, eerder as om patologie wat teenwoordig is, te bevestig. Gevolgtrekking: The LODOX is ‘n uitstekende bewysstuk in die aantoon van ‘n pneumotoraks, maar huidiglik het dit beperkte waarde as onafhanklike toets in die veld van Geregtelike Patologie. Desnieteenstaande word die verdere gebruik van LODOX as bydraende ondersoek, sowel as die prospektiewe studie van sy toepaslikheid aanbeveel.

Automated tissue image analysis aims to develop algorithms for a variety of histological applications. This has important implications in the diagnostic grading of cancer such as in breast and prostate tissue, as well as in the quantification of prognostic and predictive biomarkers that may help assess the risk of recurrence and the responsiveness of tumors to endocrine therapy. In this thesis, we use pattern recognition and image analysis techniques to solve several problems relating to histopathology and immunohistochemistry applications. In particular, we present a new method for the detection and localization of tissue microarray cores in an automated manner and compare it against conventional approaches. We also present an unsupervised method for color decomposition based on modeling the image formation process while taking into account acquisition noise. The method is unsupervised and is able to overcome the limitation of specifying absorption spectra for the stains that require separation. This is done by estimating reference colors through fitting a Gaussian mixture model trained using expectation-maximization. Another important factor in histopathology is the choice of stain, though it often goes unnoticed. Stain color combinations determine the extent of overlap between chromaticity clusters in color space, and this intrinsic overlap sets a main limitation on the performance of classification methods, regardless of their nature or complexity. In this thesis, we present a framework for optimizing the selection of histological stains in a manner that is aligned with the final objective of automation, rather than visual analysis. Immunohistochemistry can facilitate the quantification of biomarkers such as estrogen, progesterone, and the human epidermal growth factor 2 receptors, in addition to Ki-67 proteins that are associated with cell growth and proliferation. As an application, we propose a method for the identification of paired antibodies based on correlating probability maps of immunostaining patterns across adjacent tissue sections. Finally, we present a new feature descriptor for characterizing glandular structure and tissue architecture, which form an important component of Gleason and tubule-based Elston grading. The method is based on defining shape-preserving, neighborhood annuli around lumen regions and gathering quantitative and spatial data concerning the various tissue-types.

Submitted by Ricardo Cedraz Duque Moliterno (ricardo.moliterno@uefs.br) on 2016-09-13T21:44:53Z No. of bitstreams: 1 Disserta??o_George.pdf: 4996097 bytes, checksum: ece2301b72ccb1d9d33a2e2837531079 (MD5)
Made available in DSpace on 2016-09-13T21:44:53Z (GMT). No. of bitstreams: 1 Disserta??o_George.pdf: 4996097 bytes, checksum: ece2301b72ccb1d9d33a2e2837531079 (MD5) Previous issue date: 2016-02-29
Coordena??o de Aperfei?oamento de Pessoal de N?vel Superior - CAPES
The realization of an accurate diagnosis from histological images requires pathologists with practical experience because the characteristics of these images lead to a subjective analysis, which often hamper the accuracy of diagnosis. Systems that help to achieve better diagnoses can minimize doubts and improve the quality of diagnosis, influencing on increasing the effectiveness of medical treatments. This paper describes the research and development of PathoSpotter, a computer system to aid in the identification of diseases from histological images. The PathoSpotter proposes to reduce the lack of support work to histopathological diagnosis of renal diseases since much has been done in the area of cancer, but there is few published material in relation to the Digital Pathology applied to nephrology and hepatology. Our goal in this study was to apply the PathoSpotter the classification of proliferative glomerulopathy, which is a family of primary diseases affecting the kidneys. The work was based on a data set consisting of 811 histological pictures glomeruli and classical techniques of processing digital images and histopathology were used. The PathoSpotter presented a performance of 88.4% accuracy, which was similar to other Digital Pathology jobs that can be found in the literature.
A realiza??o do diagn?stico preciso a partir de imagens histol?gicas requer m?dicos patologistas com vasta experi?ncia pr?tica, pois as caracter?sticas dessas imagens conduzem a uma an?lise subjetiva que muitas vezes dificultam a exatid?o do diagn?stico. Sistemas que auxiliam a obten??o de melhores diagn?sticos podem minimizar d?vidas e melhorar a qualidade dos diagn?sticos, influenciando no aumento da efic?cia dos tratamentos m?dicos. Este trabalho descreve a pesquisa e o desenvolvimento do PathoSpotter, um sistema computacional para aux?lio na identifica??o de patologias a partir de imagens histol?gicas. O PathoSpotter se prop?e a reduzir a car?ncia de trabalhos de apoio ao diagn?stico histopatol?gico das doen?as renais, j? que muito tem sido feito na ?rea de neoplasias, mas h? pouco material publicado em rela??o ? Patologia Digital aplicada ? nefrologia ou hepatologia. Nosso objetivo neste trabalho foi aplicar o PathoSpotter na classifica??o das glomerulopatias proliferativas, que ? uma fam?lia de doen?as prim?rias que afetam os rins. O trabalho se baseou em um conjunto de dados composto por 811 imagens histol?gicas de glom?rulos, e foram utilizadas t?cnicas cl?ssicas de processamento de imagens e histopatologia digital. O PathoSpotter apresentou um desempenho de 88,4% de acur?cia, resultado similar ao de outros trabalhos de Patologia Digital que podem ser encontrados na literatura especializada.

Color theory was first formalized in the seventeenth century by Isaac Newton just a couple of decades after the first microscope was built. But it was not until the twentieth century that technological advances led to the integration of color theory, optical spectroscopy and light microscopy through spectral image processing. However, while the focus of image processing often concerns modeling of how images are perceived by humans, the goal of image processing in natural sciences and medicine is the objective analysis. This thesis is focused on color theory that promotes quantitative analysis rather than modeling how images are perceived by humans. Color and fluorescent dyes are routinely added to biological specimens visualizing features of interest. By applying spectral image processing to histopathology, subjectivity in diagnosis can be minimized, leading to a more objective basis for a course of treatment planning. Also, mathematical models for spectral image processing can be used in biotechnology research increasing accuracy and throughput, and decreasing bias. This thesis presents a model for spectral image formation that applies to both fluorescence and transmission light microscopy. The inverse model provides estimates of the relative concentration of each individual component in the observed mixture of dyes. Parameter estimation for the model is based on decoupling light intensity and spectral information. This novel spectral decomposition method consists of three steps: (1) photon and semiconductor noise modeling providing smoothing parameters, (2) image data transformation to a chromaticity plane removing  intensity variation while maintaining chromaticity differences, and (3) a piecewise linear decomposition combining advantages of spectral angle mapping and linear decomposition yielding relative dye concentrations. The methods described herein were used for evaluation of molecular biology techniques as well as for quantification and interpretation of image-based measurements. Examples of successful applications comprise quantification of colocalization, autofluorescence removal, classification of multicolor rolling circle products, and color decomposition of histological images.

Au cours de la dernière décennie, la pathologie numérique a été améliorée grâce aux avancées des algorithmes d'analyse d'images et de la puissance de calcul. Néanmoins, le diagnostic par un expert à partir d'images histopathologiques reste le gold standard pour un nombre considérable de maladies notamment le cancer. Ce type d'images préserve la structure des tissus aussi proches que possible de leur état vivant. Ainsi, cela permet de quantifier les objets biologiques et de décrire leur organisation spatiale afin de fournir une description plus précise des tissus malades. L'analyse automatique des images histopathologiques peut avoir trois objectifs: le diagnostic assisté par ordinateur, l'évaluation de la sévérité des maladies et enfin l'étude et l'interprétation des mécanismes sous-jacents des maladies et leurs impacts sur les objets biologiques. L'objectif principal de cette thèse est en premier lieu de comprendre et relever les défis associés à l'analyse automatisée des images histologiques. Ensuite, ces travaux visent à décrire les populations d'objets biologiques présents dans les images et leurs relations et interactions à l'aide des statistiques spatiales et également à évaluer la significativité de leurs différences en fonction de la maladie par des tests statistiques. Après une étape de séparation des populations d'objets biologiques basée sur la couleur des marqueurs, une extraction automatique de leurs emplacements est effectuée en fonction de leur type, qui peut être ponctuel ou surfacique. Les statistiques spatiales, basées sur la distance pour les données ponctuelles, sont étudiées et une fonction originale afin de mesurer les interactions entre deux types de données est proposée. Puisqu'il a été montré dans la littérature que la texture d'un tissu est altérée par la présence d'une maladie, les méthodes fondées sur les motifs binaires locaux sont discutées et une approche basée sur une modification de la résolution de l'image afin d'améliorer leur description est introduite. Enfin, les statistiques descriptives et déductives sont appliquées afin d'interpréter les caractéristiques extraites et d'étudier leur pouvoir discriminant dans le cadre de l'étude des modèles animaux de cancer colorectal. Ce travail préconise la mesure des associations entre différents types d'objets biologiques pour mieux comprendre et comparer les mécanismes sous-jacents des maladies et leurs impacts sur la structure des tissus. En outre, nos expériences confirment que l'information de texture joue un rôle important dans la différenciation des deux modèles d'implantation d'une même maladie
During the last decade, digital pathology has been improved thanks to the advance of image analysis algorithms and calculus power. However, the diagnosis from histopathology images by an expert remains the gold standard in a considerable number of diseases especially cancer. This type of images preserves the tissue structures as close as possible to their living state. Thus, it allows to quantify the biological objects and to describe their spatial organization in order to provide a more specific characterization of diseased tissues. The automated analysis of histopathological images can have three objectives: computer-aided diagnosis, disease grading, and the study and interpretation of the underlying disease mechanisms and their impact on biological objects. The main goal of this dissertation is first to understand and address the challenges associated with the automated analysis of histology images. Then it aims at describing the populations of biological objects present in histology images and their relationships using spatial statistics and also at assessing the significance of their differences according to the disease through statistical tests. After a color-based separation of the biological object populations, an automated extraction of their locations is performed according to their types, which can be point or areal data. Distance-based spatial statistics for point data are reviewed and an original function to measure the interactions between point and areal data is proposed. Since it has been shown that the tissue texture is altered by the presence of a disease, local binary patterns methods are discussed and an approach based on a modification of the image resolution to enhance their description is introduced. Finally, descriptive and inferential statistics are applied in order to interpret the extracted features and to study their discriminative power in the application context of animal models of colorectal cancer. This work advocates the measure of associations between different types of biological objects to better understand and compare the underlying mechanisms of diseases and their impact on the tissue structure. Besides, our experiments confirm that the texture information plays an important part in the differentiation of two implemented models of the same disease

Au cours de la dernière décennie, la pathologie numérique a été améliorée grâce aux avancées des algorithmes d'analyse d'images et de la puissance de calcul. Néanmoins, le diagnostic par un expert à partir d'images histopathologiques reste le gold standard pour un nombre considérable de maladies notamment le cancer. Ce type d'images préserve la structure des tissus aussi proches que possible de leur état vivant. Ainsi, cela permet de quantifier les objets biologiques et de décrire leur organisation spatiale afin de fournir une description plus précise des tissus malades. L'analyse automatique des images histopathologiques peut avoir trois objectifs: le diagnostic assisté par ordinateur, l'évaluation de la sévérité des maladies et enfin l'étude et l'interprétation des mécanismes sous-jacents des maladies et leurs impacts sur les objets biologiques. L'objectif principal de cette thèse est en premier lieu de comprendre et relever les défis associés à l'analyse automatisée des images histologiques. Ensuite, ces travaux visent à décrire les populations d'objets biologiques présents dans les images et leurs relations et interactions à l'aide des statistiques spatiales et également à évaluer la significativité de leurs différences en fonction de la maladie par des tests statistiques. Après une étape de séparation des populations d'objets biologiques basée sur la couleur des marqueurs, une extraction automatique de leurs emplacements est effectuée en fonction de leur type, qui peut être ponctuel ou surfacique. Les statistiques spatiales, basées sur la distance pour les données ponctuelles, sont étudiées et une fonction originale afin de mesurer les interactions entre deux types de données est proposée. Puisqu'il a été montré dans la littérature que la texture d'un tissu est altérée par la présence d'une maladie, les méthodes fondées sur les motifs binaires locaux sont discutées et une approche basée sur une modification de la résolution de l'image afin d'améliorer leur description est introduite. Enfin, les statistiques descriptives et déductives sont appliquées afin d'interpréter les caractéristiques extraites et d'étudier leur pouvoir discriminant dans le cadre de l'étude des modèles animaux de cancer colorectal. Ce travail préconise la mesure des associations entre différents types d'objets biologiques pour mieux comprendre et comparer les mécanismes sous-jacents des maladies et leurs impacts sur la structure des tissus. En outre, nos expériences confirment que l'information de texture joue un rôle important dans la différenciation des deux modèles d'implantation d'une même maladie
During the last decade, digital pathology has been improved thanks to the advance of image analysis algorithms and calculus power. However, the diagnosis from histopathology images by an expert remains the gold standard in a considerable number of diseases especially cancer. This type of images preserves the tissue structures as close as possible to their living state. Thus, it allows to quantify the biological objects and to describe their spatial organization in order to provide a more specific characterization of diseased tissues. The automated analysis of histopathological images can have three objectives: computer-aided diagnosis, disease grading, and the study and interpretation of the underlying disease mechanisms and their impact on biological objects. The main goal of this dissertation is first to understand and address the challenges associated with the automated analysis of histology images. Then it aims at describing the populations of biological objects present in histology images and their relationships using spatial statistics and also at assessing the significance of their differences according to the disease through statistical tests. After a color-based separation of the biological object populations, an automated extraction of their locations is performed according to their types, which can be point or areal data. Distance-based spatial statistics for point data are reviewed and an original function to measure the interactions between point and areal data is proposed. Since it has been shown that the tissue texture is altered by the presence of a disease, local binary patterns methods are discussed and an approach based on a modification of the image resolution to enhance their description is introduced. Finally, descriptive and inferential statistics are applied in order to interpret the extracted features and to study their discriminative power in the application context of animal models of colorectal cancer. This work advocates the measure of associations between different types of biological objects to better understand and compare the underlying mechanisms of diseases and their impact on the tissue structure. Besides, our experiments confirm that the texture information plays an important part in the differentiation of two implemented models of the same disease

Tese de mestrado integrado, Engenharia Biomédica e Biofísica (Radiações em Diagnóstico e Terapia) Universidade de Lisboa, Faculdade de Ciências, 2019
O cancro de próstata é o segundo tipo de cancro com maior prevalência nos homens, em todo o mundo. A deteção inicial desta doença ocorre, geralmente, durante exames e consultas de rotina, quando níveis aumentados do antigénio prostático específico e/ou um exame retal anormal são descobertos. Contudo, apenas a avaliação histopatológica, inicialmente baseada em amostras extraídas através de uma biópsia, é capaz de fornecer um diagnóstico definitivo, permitindo não só orientar o tratamento do doente, assim como o processo de tomada de decisão associado. Após esta avaliação inicial, se o doente for diagnosticado com cancro da próstata localizado, ou seja, um tumor confinado à próstata, o tratamento mais adotado é a prostatectomia radical. Este último é o tratamento padrão utilizado quando se pretende uma terapia com curativa. A vantagem da técnica de prostatectomia radical é que, toda a próstata, onde o tumor se encontra confinado, é removida cirurgicamente, auxiliando na redução do risco de metástases. A posterior análise da peça prostática permite ao patologista avaliar diversas características tumorais, determinantes no prognóstico do doente. Para este fim, o microscópio tem sido a principal ferramenta utilizada, uma vez que proporciona imagens ao vivo com uma ótima resolução. No entanto, desde a introdução do primeiro sistema automatizado de digitalização de lâminas em imagens de alta resolução, o interesse da comunidade de anatomia patológica em explorar este tipo de métodos para diferentes aplicações tem crescido exponencialmente. O potencial desta nova área não é, todavia, a simples transferência de uma imagem da lâmina de vidro para um monitor, nem tão pouco a flexibilidade de distribuição e modificação da própria imagem digital, mas sim, a possibilidade de aprimorar a avaliação do patologista com informações e inteligência que não podem ser detetados pela análise humana. Por conseguinte, a implementação de algoritmos de aprendizagem automática capazes de executar tarefas como deteção, classificação e segmentação de imagens digitais histopatológicas é, finalmente, possível. Estes métodos de análise automatizada permitem explorar todo o panorama morfológico do tumor e dos seus elementos mais invasivos presentes, capturando, por exemplo, a orientação nuclear, a textura, a forma e a arquitetura. A complexidade e densidade inerente a este tipo de imagens, oferecem uma abundância de informação, ideal para estimular e promover o desenvolvimento de algoritmos baseados em deep learning. No que diz respeito ao cancro da próstata, os algoritmos desenvolvidos visam apoiar as avaliações efetuados pelos patologistas, nomeadamente, estadiamento e classificação, sendo, portanto, focados no sistema de classificação utilizada para a próstata - o Gleason score. Contudo, problemáticas alternativas podem também beneficiar da aplicação destas t´técnicas, nomeadamente métodos capazes de distinguir imagens que apenas contêm tecido benigno de imagens em que tumor esteja presente. Por outro lado, modelos capazes de prever a recidiva bioquímica de cancro da próstata permitiriam aos médicos modificar estratégias de tratamento e pós-tratamento, a fim de equilibrar benefícios e efeitos adversos de um terapia específica. A previsão de recidiva permite, também, que os pacientes escolham com responsabilidade os diferentes tratamentos e estratégias que lhes são propostos pelos médicos, possibilitando, em última análise, uma maior sua satisfação após o tratamento. Desta forma, com o objetivo de explorar os referidos problemas, a presente dissertação apresenta procedimentos de recolha, processamento e anotação de dados, que permitiram a criação de uma base digital de dados histológicos anotados da próstata. Com base nestes dados dois modelos distintos de deep learning, especificamente Convolutional Neural Networks foram desenvolvidos. O modelo I propõe a identificação de cancro da próstata e diferenciação entre tumor e tecido benigno. O modelo II pretende prever a condição de recidiva bioquímica do cancro de próstata, para um período de tempo posterior `a cirurgia em dois anos. Relativamente ao desenvolvimento da base de dados, 200 casos de cancro da próstata, tratados através de prostatectomia radical, foram selecionados. As lâminas correspondentes à lesão índice, ou seja, à lesão principal, foram identificadas e, apenas estas foram incluídas na amostra final. A adoção desta abordagem deveu-se ao facto de que cada peça origina entre 15-45 lâminas, sendo que a maioria não contém tumor. Por outro lado, dado o período de tempo para a realização de todo este projeto, seria inviável a utilização de todas as lâminas. Assim, as lâminas selecionadas foram digitalizadas e processadas. Uma t´técnica de normalização de contraste foi aplicada, de forma a uniformizar as cores das diferentes imagens digitais, evitando uma elevada variabilidade de contraste e cor que advém da utilização de diferentes protocolos de cor, bem como da própria digitalização da lâmina. As imagens histológicas digitalizadas e normalizadas foram posteriormente divididas em imagens mais pequenas, isto é, subimagens, uma vez que desta forma existe uma otimização da extração de características por parte dos algoritmos. Estas subimagens foram individualmente visualizadas e anotadas, originando um total de cerca de 160,000 subimagens, correspondentes aos 200 casos diferentes selecionados. Para o desenvolvimento do modelo de classificação do cancro de próstata, a arquitetura Inception v3 foi implementada e treinada utilizando as subimagens da base de dados. Este modelo foi capaz de identificar três classes distintas: negativa (tecido benigno), positiva (tecido maligno) e neoplasia intraepitelial da próstata, esta última, embora com menor precisão dada a quantidade reduzida de exemplos pertencentes a esta classe. Um valor de 93 % de precisão foi obtido, o que corresponde a valores equiparados ao estado da arte para este tipo de técnicas. Este valor, contudo, demonstra ainda potencial para otimização e melhoria, uma vez que as diferentes classes dos dados utilizados seguiam uma distribuição não equilibrada. A inclusão de mais casos clínicos e a aplicação de técnicas de aumento de dados, podem ser facilmente realizadas, o que culminará num modelo com ainda melhor precisão de classificação. Relativamente ao modelo referente à previsão de recidiva bioquímica, a mesma arquitetura foi utilizada, mas neste caso, treinada apenas com base nas subimagens positivas, isto ´e, as subimagens contendo tecido maligno da próstata. Os resultados obtidos revelaram que este modelo não tem a capacidade de extrair informação relevante correlacionada com o objetivo do estudo, e portanto, não consegue distinguir com sucesso casos não recorrentes de casos recorrentes, produzindo apenas uma precisão de 60 %. Contudo, apesar do referido modelo falhar na execução do objetivo estipulado, é fundamental notar que a tarefa de predição de recidiva bioquímica é de complexidade elevada, não sendo possível aos patologistas, através da observação das imagens histológicas, retirar nenhuma conclusão que diretamente se correlacione com esta condição. Diferentes abordagens, como por exemplo, o aumento da quantidade de dados utilizados, a introdução no modelo de características clínicas relevantes no prognóstico da doença poderão apresentar melhorias substanciais, no que diz respeito à capacidade preditiva deste modelo. Concluindo, a capacidade de algoritmos de deep learning para extrair informação relevante de imagens digitais da histopatologia da próstata foi demonstrada através do presente estudo. O desenvolvimento e o criação de uma base de dados anotados, fornece a base fundamental para o desenvolvimento de modelos adicionais, onde diversas questões podem ser exploradas. O desenvolvimento de uma interface que permita implementar o modelo de deteção de cancro da próstata desenvolvido é também uma possibilidade, uma vez que fornece eficiência e consistência, beneficiando a prática da patologia clinica.
Prostate cancer is the second most prevalent cancer in men, worldwide. Histopathological assessment plays an indispensable role in understanding the disease mechanisms, providing definitive diagnosis to guide patient treatment and management decisions. The microscope has been the primary tool to this end, producing images at increased resolution. However, with the development of the first automated high resolution whole-slide imaging system, which allows the digitisation of glass slides, interest in using this system for different applications in pathology practice has steadily grown, giving rise to the digital pathology field. The promise of digital pathology is not, however, the simple transfer of an image from a glass slide to a monitor, not even the flexibility of distribution and modification of the image, but instead the potential to enhance the pathologist’s assessment with information and artificial intelligence that cannot be perceived by human examination. With the advent of digital pathology and the recent expert-level accuracy achieved by machine learning based algorithms in medical image detection, classification and segmentation, new possibilities to develop automated image analysis methods arise. As far as prostate cancer is concerned, these models have been aimed at supporting pathologist’s image descriptions such as staging and grading, being hence, focused on the Gleason grading system. In order to explore alternative problems, the present dissertation presents data collection, processing and annotation procedures, that allowed the creation of an annotated digital histology database of prostate resection cases. These data was used to develop deep learning models not only to classify prostate cancer, but also to predict prostate cancer biochemical recurrence. Inception v3 architecture was implemented and trained from scratch for the proposed assignments. The prostate cancer classification model yielded an accuracy of 93%, being able to identify three distinct classes: negative (benign tissue), positive (malignant tissue) and high-grade prostate intraepithelium neoplasia, the latter, although, with lower precision, given to unbalanced class distributions. The prostate cancer biochemical prediction model was not able to successfully distinguish between non-recurrent and recurrent cases, yielding an accuracy of 60%. This value was accomplished, nevertheless, due to the fact that the model was classifying all entries as negative, and therefore, the value of accuracy corresponds to the percentage of negative cases present in the dataset. Although not all models here developed achieved good results, the capacity of deep learning algorithms to harvest relevant features from prostate histopathology digital images has been demonstrated. The development and establishment of an annotated database provides the fundamental basis to further develop additional models, and mainly to improve the biochemical recurrence prediction model by applying more sophisticated methods, given the complexity of this problem.

INTRODUCTION: Ebola virus (EBOV) is a highly infectious and often lethal filovirus that causes hemorrhagic fever, with a reported case fatality rate of 40-90%. There are currently no Food and Drug Administration (FDA) approved medical countermeasures (MCMs) for EBOV. Non-human primates (NHPs) remain the gold standard animal model for EBOV research as they most accurately recapitulate human disease. OBJECTIVE: This study aimed to characterize the temporal viral pathogenesis of EBOV in the liver of infected rhesus macaques using routine histopathology, multiplex immunohistochemistry (mIHC) and multiplex fluorescent In Situ Hybridization (mFISH), refined by digital pathology (DP) and image analysis (DIA). METHODS: 21 FFPE liver sections from EBOV-infected rhesus macaques were examined microscopically (Uninfected controls n=3; 3 DPE n=3; 4 DPE n=3; 5 DPE n=3; 6 DPE n=3; Terminal n=6). Tissues were stained with H&E and PTAH for histopathological scoring. Three serial sections were fluorescently immunolabeled or hybridized under three independent conditions (1.EBOV VP35, Tissue Factor, CD68; 2.EBOV VP35, Heppar, Myeloperoxidase (MPO); 3.EBOV VP35, IL-6, ISG-15). Slides were digitized by a Vectra PolarisTM fluorescent whole slide scanner and DIA was conducted using HaloTM image analysis software. Statistical analysis was conducted using GraphPad PrismTM 8.0. RESULTS: Comparing peracute (3-4 DPE) to acute (5-6 DPE) and terminal (6-8 DPE) EBOV infection, there is a statistically significant (p < 0.05) increase in hepatic inflammation and fibrin thrombi, correlating with an absolute increase in macrophages (CD68), neutrophils (MPO), and total % of Tissue Factor in the liver. There is also a significant increase in the severity of necrosis, which correlates with a decrease in Heppar. While there was significant colocalization of VP35 and CD68 starting at 4 DPE, there was only rare colocalization of VP35 with Heppar, even in terminal animals. Similar to mIHC, progressive and statistically significant differences were observed in gene expression when comparing peracute to acute and terminal EBOV infection. IL-6 predominated within periportal fibrovascular compartments, but also colocalized within cells concurrently expressing EBOV VP35. EBOV VP35 expression was observed within histiocytes, endothelial cells, and less commonly hepatocytes. ISG-15 expression was observed in periportal regions and in proximity to cells expressing EBOV VP35, but colocalization within EBOV VP35 expressing cells was an extremely rare event. CONCLUSION: Qualitative tools are well suited for confirming virulence and viral tissue tropism, but do little to build on our current understanding of disease. Using DIA in partnership with mIHC and mFISH, this study quantified statistically significant temporal changes in the immunoreactivity and hybridization of host and viral biomarkers that have previously been linked to the pathogenesis of EBOV. Taken together, these tools have enabled us to characterize minute changes that reflect magnitudes of biological variability simply not feasible to detect with the human eye. Furthermore, spatial context has refined our current understanding of differential gene expression of EBOV, which has the potential to aid in development of host-directed therapies. The establishment of these benchmarks will serve as a guide for the validation of cross-institutional EBOV animal models.