Abstract:

This talk will summarize our recent work on using deep learning for medical image analysis. Deep learning is a power tool that can discover new features suitable for different applications. Although the conventional human-made filters can be used to extract certain advanced features, it is time-consuming to discover a new filter and also the extracted features may not necessarily fit a particular study under consideration. Besides, a lot of efforts need to spend on the testing and selection of different choices of human-made features, which is difficult for the researchers with limited experience to select suitable features. On the other hand, deep learning is designed to automatically discover features, from a set of given data, for each particular application. Thus, it is able to discover new features that were never discovered by researchers before. Since 2012, we started to apply deep learning for various medical image analysis applications, e.g., image segmentation, registration, and disease classification, all of which can be formulated as feature-matching problems and thus can be solved effectively with the learned new features by deep learning. In this talk, I will demonstrate the applications of deep learning in segmenting hippocampus, registering brain images, and identifying brain disorders from multi-modality data (in the field of neuroimaging). I will also show the results on segmenting prostate from MR images, which is important for in vivo diagnosis of prostate cancer and also the radiotherapy of prostate cancer.

2) Gang Li, Image Analysis Core Lab, l’'Université de Caroline du Nord à Chapel Hill, NC, USA

Titre : Apprentissage profond pour la classification d'images MR

Dans l'objectif de construire un processus de traitement automatique d'images médicales, la première étape est d'identifier la région anatomique correspondante à une image, afin de pouvoir la rediriger vers des algorithmes spécialisés de segmentation. Nous utilisons pour cela des méthodes d'apprentissage profond, qui bien que donnant d'excellents résultats, soulèvent un certain nombre de questions. Comment choisir l'architecture du réseau ? Comment analyser les résultats et les améliorer ?

Présentations courtes :

1) Automatic segmentation and characterization of brain tumors using robust multivariate

clustering of multiparametric MRI

Alexis Arnaud 1,2, Florence Forbes 1,2, Emmanuel L. Barbier 3,4, et Benjamin Lemasson 3,4 1 INRIA, 2 LJK Université Grenoble Alpes, 3 U836 INSERM, 4 GIN Université Grenoble Alpes

Alexis Arnaud

Abstract : Brain tumor segmentation is a difficult task in the field of multiparametric MRI analysis because of the number of maps that are available. Furthermore, the characterization of brain tumors can be timeconsuming, even for medical experts, and the reference method is biopsy which is a local and invasive technique. Because of this, it is important to develop automatic and non-invasive approaches in order to help the medical expert with these issues. In this study we use a robust statistical model-based method to classify multiparametric MRI of rat brains. The voxels are gather into classes resulting from Multivariate Multi-scaled Student distributions, which can accommodate outliers. First we adjust a mixture model on a reference group of rats to learn the MRI characteristics of healthy tissues. Second we use this model to delineate the brain tumors as atypical voxels in the data set of unhealthy rats. Third we adjust a new mixture model only on the atypical voxels to learn the MRI characteristics of tumorous tissues. Finally, we extract a fingerprint for each tumor type to make a tumor dictionary. Our data set is composed of healthy rats (n=8 rats) and 4 groups of rats bearing a brain tumor model (n=8 per group). For each rat, we acquired 5 quantitative MRI parameters along 5 slices. And the proposed tumor dictionary reaches a rate of 75% of accurate prediction with a leave-one-out procedure.

2) Dictionary Learning for Pattern Classification in Medical Imaging

Hrishikesh Deshpande, Christian Barillot,Pierre Maurel, INRIA Rennes,

Abstract: Most natural signals can be approximated by a linear combination of a few atoms in a dictionary. Such sparse representations of signals and dictionary learning (DL) methods have received a special attention over the past few years. While standard DL approaches are effective in applications such as image denoising or compression, several discriminative DL methods have been proposed to achieve better image classification. In this talk, I will present our recent work on dictionary learning for pattern classification in medical imaging. This work demonstrated that the dictionary size for each class is an important factor in the pattern recognition applications where there exist variability difference between classes, in the case of both the standard and discriminative DL methods. The proposition of using different dictionary size based on complexity of the class data was validated in a computer vision application such as lips detection in face images, followed by more complex medical imaging application such as classification of multiple sclerosis (MS) lesions using MR images. The class specific dictionaries were learned for the MS lesions and individual healthy brain tissues, and the size of the dictionary for each class is adapted according to the complexity of the underlying data. The algorithm is validated using 52 multi-sequence MR images acquired from 13 MS patients.

3) Scale-Adaptive Supervoxel-based Random Forests for Liver Tumor Segmentation in Dynamic Contrast-Enhanced CT Scans

Pierre-Henri Conze? , Vincent Noblet? , François Rousseau** , Fabrice Heitz? , Vito de BlasiM, Riccardo Memeo *** and Patrick Pessaux ***

Pierre-Henri Conze conze@unistra.fr

Affiliations: ICube UMR 7357, Université de Strasbourg, CNRS, Federation de Médecine Translationnelle de Strasbourg. * Institut Mines-Telecom, Telecom Bretagne, INSERM, LATIM UMR 1101. *** Department of Hepato-Biliary and Pancreatic Surgery, Nouvel Hospital Civil, Institut Hospitalo-Universitaire de Strasbourg.

Abstract --Towards an efficient clinical management of hepato-cellular carcinoma (HCC), we propose a multi-label tissue classification framework dedicated to tumor necrosis rate estimation from dynamic contrast-enhanced CT scans. Based on machine learning, it requires weak interaction efforts to segment healthy, active and necrotic liver tissues. Our contributions are two-fold. First, we apply random forest (RF) on supervoxels using multi-phase supervoxel-based features that discriminate tissues based on their dynamic in response to contrast agent injection. Second, we extend this technique in a hierarchical multi-scale fashion to deal with multiple spatial extents and appearance heterogeneity. It translates in an adaptive data sampling scheme combining RF and hierarchical multi-scale tree resulting from recursive supervoxel decomposition. By concatenating multi-phase features across the hierarchical multi-scale tree to describe leaf supervoxels, we enable RF to automatically infer the most informative scales without defining any explicit rules on how to combine them. Assessment on clinical data confirms the benefits of multi-phase information embedded in a multi-scale supervoxel representation for HCC tumor segmentation.

5) Bayesian Inference for Biomarker Discovery in Proteomics

Audrey Giremus et Jean-François Giovannelli. Université de Bordeaux

Giremus

Abstract -- The presented work addresses the question of biomarker discovery in proteomics. More precisely, for a set of individuals, a status (Healthy or Pathological) and the concentrations for a given list of proteins are available. The tackled problem is to extract a short sub-list of proteins, namely the biomarkers, that enables to model the status. The work presents two cases. The first one accounts for biological variabilities and it is founded on natural models (Gaussian for the concentrations and Bernoulli for the status). It does not impose constraints in terms of a regression model. The second one includes in addition technological variabilities that may significantly impact observed concentrations. They are referred to as noiseless and noisy models, respectively. The developed selection strategy for both models is optimal in the sense that it minimizes a global mean error (misdetection and false identification). It is developed in a Bayesian framework and practically it amounts to selecting the model with the higher posterior probability. The key difficulty is to calculate these probabilities since they are based on the evidences that require marginalization. The key point of the work is that: for the noiseless case, we demonstrate the analytical solutions and for the noisy case, we propose an approximated solution. The methods are numerically assessed and compared to two existing methods on synthetic and clinical data.

6) Cancer Therapy Outcome Prediction based on Dempster-Shafer Theory and PET Imaging

Chunfeng Lian, Su Ruan, Thierry Denoeux, Pierre vera, Heudiasyc, UTC & LITIS Université de Rouen

Chunfeng Lian, Chunfeng Lian chunfeng.lian@gmail.com

Abstract: In cancer therapy, utilizing FDG-PET image-based features for accurate outcome prediction is challenging because of 1) limited discriminative information within a small number of PET image sets, and 2) fluctuant feature characteristics caused by the inferior spatial resolution and system noise of PET imaging. In this study, we proposed a Dempster-Shafer theory (DST) based approach to accurately predict cancer therapy outcome with both PET imaging features and clinical characteristics. First, a specific loss function with sparse penalty was developed to learn an adaptive low-rank distance metric for representing the dissimilarity between different patients' feature vectors. By minimizing this loss function, a linear low-dimensional transformation of input features was achieved. Also, imprecise features were excluded simultaneously by applying a sparsity regularization of the learnt dissimilarity metric in the loss function. Finally, the learnt dissimilarity metric was applied in an evidential K-nearest-neighbor (EK-NN) classifier to predict treatment outcome. The proposed method has been evaluated by two clinical datasets, showing good performance.

7) Un modèle de mélange croisé Poissonien pour la classification efficace de données médicales textuelles

Melissa Ailem, François Role, Mohamed Nadif , Université Paris Descartes

{melissa.ailem,francois.role,mohamed.nadif}@parisdescartes.fr LIPADE

Résumé : Au cours de la dernière décennie, plusieurs études ont démontré l?importance de la classification croisée pour produire simultanément des groupes d?objets et d?attributs. Même si l?objectif est d?obtenir une classification sur une seule dimension, les techniques de classification croisée sont souvent plus efficaces que les techniques de classification simples, en particulier lorsque l?on considère des données creuses (sparses) à grande dimension. Dans ce travail, nous proposons un modèle de blocs latents parcimonieux avec des distributions de Poisson sous contraintes (Sparse Poisson Latent Block Model (SPLBM)). Un point important est que ce modèle comprend la sparsité dans sa formulation, ce qui le rend particulièrement adapté à la classification croisée de matrices document-terme. Les avantages du modèle proposé sont de deux ordres. Tout d?abord, c?est un modèle probabiliste rigoureux et très parcimonieux. Deuxièmement, il a été spécialement conçu pour faire face aux problèmes de sparsité des données. En conséquence, en plus de la recherche de blocs homogènes, comme les autres algorithmes disponibles, il filtre les blocs non-informatifs qui sont dus à la sparsité des données. Des expériences sur une matrice document-terme biomédicale construite à partir de la base de données PUBMED et contenant des documents sur cinq maladies différentes, montrent l?efficacité de l?algorithme proposé pour révéler des co-clusters pertinents, co-clusters qui sont clairement indicatifs de chaque maladie. En outre, d?autres expériences sur diverses données textuelles provenant de différents domaines, de taille et de structure différentes montrent qu?un algorithme basé sur SPLBM surpasse clairement l?état de l?art des algorithmes de classification croisée. Plus particulièrement, l?algorithme proposé réussit à intercepter les structures naturelles des co-clusters issus de jeux de données asymétriques, des structures que d?autres algorithmes connus sont incapables de traiter efficacement.