In this paper, we propose a new theoretical analysis of unsupervised domain adaptation that relates notions of large margin separation, adversarial learning and optimal transport. This analysis generalizes previous work on the subject by providing a bound on the target margin violation rate, thus reflecting a better control of the quality of separation between classes in the target domain than bounding the misclassification rate. The bound also highlights the benefit of a large margin separation on the source domain for adaptation and introduces an optimal transport (OT) based distance between domains that has the virtue of being task-dependent, contrary to other approaches. From the obtained theoretical results, we derive a novel algorithmic solution for domain adaptation that introduces a novel shallow OT-based adversarial approach and outperforms other OT-based DA baselines on several simulated and real-world classification tasks.

Delineation of the cardiac structures from 2D echocardiographic images is a common clinical task to establish a diagnosis. Over the past decades, the automation of this task has been the subject of intense research. In this paper, we evaluate how far the state-of-the-art encoder-decoder deep convolutional neural network methods can go at assessing 2D echocardiographic images, i.e., segmenting cardiac structures and estimating clinical indices, on a dataset, especially, designed to answer this objective. We, therefore, introduce the cardiac acquisitions for multi-structure ultrasound segmentation dataset, the largest publicly-available and fully-annotated dataset for the purpose of echocardiographic assessment. The dataset contains two and four-chamber acquisitions from 500 patients with reference measurements from one cardiologist on the full dataset and from three cardiologists on a fold of 50 patients. Results show that encoder-decoder-based architectures outperform state-of-the-art non-deep learning methods and faithfully reproduce the expert analysis for the end-diastolic and end-systolic left ventricular volumes, with a mean correlation of 0.95 and an absolute mean error of 9.5 ml. Concerning the ejection fraction of the left ventricle, results are more contrasted with a mean correlation coefficient of 0.80 and an absolute mean error of 5.6%. Although these results are below the inter-observer scores, they remain slightly worse than the intra-observer's ones. Based on this observation, areas for improvement are defined, which open the door for accurate and fully-automatic analysis of 2D echocardiographic images.

This paper aims at presenting results of a computer-aided diagnostic (CAD) system for voxel based detection and characterization of prostate cancer in the peripheral zone based on multiparametric magnetic resonance (mp-MR) imaging. We propose an original scheme with the combination of a feature extraction step based on a sparse dictionary learning (DL) method and a supervised classification in order to discriminate normal {N}, normal but suspect {NS} tissues as well as different classes of cancer tissue whose aggressiveness is characterized by the Gleason score ranging from 6 {GL6} to 9 {GL9}. We compare the classification performance of two supervised methods, the linear support vector machine (SVM) and the logistic regression (LR) classifiers in a binary classification task. Classification performances were evaluated over an mp-MR image database of 35 patients where each voxel was labeled, based on a ground truth, by an expert radiologist. Results show that the proposed method in addition to being explicable thanks to the sparse representation of the voxels compares well (AUC>0.8) with recent state-of-the-art performances. Preliminary visual analysis of example patient cancer probability maps indicate that cancer probabilities tend to increase as a function of the Gleason score.

We propose a new computer-aided detection scheme for prostate cancer screening on multiparametric magnetic resonance (mp-MR) images. Based on an annotated training database of mp-MR images from thirty patients, we train a novel support vector machine (SVM)-inspired classifier which simultaneously learns an optimal linear discriminant and a subset of predictor variables (or features) that are most relevant to the classification task, while promoting spatial smoothness of the malignancy prediction maps. The approach uses a ℓ1-norm in the regularization term of the optimization problem that rewards sparsity. Spatial smoothness is promoted via an additional cost term that encodes the spatial neighborhood of the voxels, to avoid noisy prediction maps. Experimental comparisons of the proposed ℓ1-Smooth SVM scheme to the regular ℓ2-SVM scheme demonstrate a clear visual and numerical gain on our clinical dataset.

Building an accurate training database is challenging in supervised classification. For instance, in medical imaging, radiologists often delineate malignant and benign tissues without access to the histological ground truth, leading to uncertain data sets. This paper addresses the pattern classification problem arising when available target data include some uncertainty information. Target data considered here are both qualitative (a class label) or quantitative (an estimation of the posterior probability). In this context, usual discriminative methods, such as the support vector machine (SVM), fail either to learn a robust classifier or to predict accurate probability estimates. We generalize the regular SVM by introducing a new formulation of the learning problem to take into account class labels as well as class probability estimates. This original reformulation into a probabilistic SVM (P-SVM) can be efficiently solved by adapting existing flexible SVM solvers. Furthermore, this framework allows deriving a unique learned prediction function for both decision and posterior probability estimation providing qualitative and quantitative predictions. The method is first tested on synthetic data sets to evaluate its properties as compared with the classical SVM and fuzzy-SVM. It is then evaluated on a clinical data set of multiparametric prostate magnetic resonance images to assess its performances in discriminating benign from malignant tissues. P-SVM is shown to outperform classical SVM as well as the fuzzy-SVM in terms of probability predictions and classification performances, and demonstrates its potential for the design of an efficient computer-aided decision system for prostate cancer diagnosis based on multiparametric magnetic resonance (MR) imaging.

We have designed a computer-aided diagnosis system to discriminate between hypermetabolic cancer lesions and hypermetabolic inflammatory or physiological but noncancerous processes in FDG PET/CT exams of lymphoma patients. Detection performance of the support vector machine (SVM) classifier was assessed based on feature sets including 105 positron emission tomography (PET) and Computed tomography (CT) characteristics derived from the clinical practice and from more sophisticated texture analysis. An original feature selection method based on combining different filter methods was proposed. The evaluation database consisted of 156 lymphomatous and 32 suspicious but nonlymphomatous regions of interest. Different types of training databases including either the PET and CT features or the PET features only, with or without feature selection, were evaluated to assess the added value of multimodality and texture information on classification performance. An optimization study was conducted for each classifier separately to select the best combination of parameters. Promising classification performance was achieved by the SVM classifier combined with the 12 most discriminant PET and CT features with a value of the area under the receiver operating curve of 0.91.

This paper presents the Virtual Imaging Platform (VIP), a platform accessible at http://vip.creatis.insa-lyon.fr to facilitate the sharing of object models and medical image simulators, and to provide access to distributed computing and storage resources. A complete overview is presented, describing the ontologies designed to share models in a common repository, the workflow template used to integrate simulators, and the tools and strategies used to exploit computing and storage resources. Simulation results obtained in four image modalities and with different models show that VIP is versatile and robust enough to support large simulations. The platform currently has 200 registered users who consumed 33 years of CPU time in 2011.

Medical image simulation produces virtual images from software representations of imaging devices and virtual object models representing the human body. Object models consist of the geometry of the objects (e.g. organs, tissues, pathological structures, etc.) and of their physical parameters used for the simulation. The diversity of this information makes it difficult to share and reuse across simulation modalities and users. We address this issue by explicitly describing object models using a semantic approach. In particular, we developed an ontology that contains the relevant concepts and relations of the domain of object models for image simulation. This ontology is used to annotate object models and to describe their content and structure. In this paper, we present the construction steps of this ontology, the representation choices and we illustrate how it is used to annotate object models.

This paper describes a framework for the integration of medical image simulators in the Virtual Imaging Platform (VIP). Simulation is widely involved in medical imaging but its availability is hampered by the heterogeneity of software interfaces and the required amount of computing power. To address this, VIP defines a simulation workflow template which transforms object models from the IntermediAte Model Format (IAMF) into native simulator formats and parallelizes the simulation computation. Format conversions, geometrical scene definition and physical parameter generation are covered. The core simulator executables are directly embedded in the simulation workflow, enabling data parallelism exploitation without modifying the simulator. The template is instantiated on simulators of the four main medical imaging modalities, namely Positron Emission Tomography, Ultrasound imaging, Magnetic Resonance Imaging and Computed Tomography. Simulation examples and performance results on the European Grid Infrastructure are shown.

Two groups of bootstrap methods have been proposed to estimate the statistical properties of positron emission tomography (PET) images by generating multiple statistically equivalent data sets from few data samples. The first group generates resampled data based on a parametric approach assuming that data from which resampling is performed follows a Poisson distribution while the second group consists of nonparametric approaches. These methods either require a unique original sample or a series of statistically equivalent data that can be list-mode files or sinograms. Previous reports regarding these bootstrap approaches suggest different results. This work compares the accuracy of three of these bootstrap methods for 3-D PET imaging based on simulated data. Two methods are based on a unique file, namely a list-mode based nonparametric (LMNP) method and a sinogram based parametric (SP) method. The third method is a sinogram-based nonparametric (SNP) method. Another original method (extended LMNP) was also investigated, which is an extension of the LMNP methods based on deriving a resampled list-mode file by drawings events from multiple original list-mode files. Our comparison is based on the analysis of the statistical moments estimated on the repeated and resampled data. This includes the probability density function and the moments of order 1 and 2. Results show that the two methods based on multiple original data (SNP and extended LMNP) are the only methods that correctly estimate the statistical parameters. Performances of the LMNP and SP methods are variable. Simulated data used in this study were characterized by a high noise level. Differences among the tested strategies might be reduced with clinical data sets with lower noise.

The purpose of the work described in this paper was the development of a framework for the creation of a realistic positron emission tomography (PET) simulated database incorporating patient-specific variability. The ground truth used was therefore based on clinical PET/computed tomography (CT) data of oncology patients. In the first step, the NURBS-based cardiac-torso phantom was adapted to the patient's CT acquisitions to reproduce their specific anatomy while the corresponding PET acquisitions were used to derive the activity distribution of each organ of interest. Secondly, realistic tumor shapes with homogeneous or heterogeneous activity distributions were modeled based on segmentation of the PET tumor volume and incorporated in the patient-specific models obtained at the first step. Lastly, patient-specific respiratory motion was also modeled. The derived patient-specific models were subsequently combined with the PET SORTEO Monte Carlo simulation tool for the simulation of the whole-body PET acquisition process. The accuracy of the simulated datasets was assessed in comparison to the original clinical patient images. In addition, a couple of applications for such simulated images were also demonstrated. Future work will focus on the creation of a comprehensive database of simulated raw data and reconstructed whole-body images, facilitating the rigorous evaluation of image-processing algorithms in PET for oncology applications.