Search

International audience; Frontotemporal dementia and amyotrophic lateral sclerosis are rare neurodegenerative diseases with no effective treatment. The development of biomarkers allowing an accurate assessment of disease progression is crucial for evaluating new therapies. Concretely, neuroimaging and transcriptomic (microRNA) data have been shown useful in tracking their progression. However, no single biomarker can accurately measure progression in these complex diseases. Additionally, large samples are not available for such rare disorders. It is thus essential to develop methods that can model disease progression by combining multiple biomarkers from small samples. In this paper, we propose a new framework for computing a disease progression score (DPS) from cross-sectional multimodal data. Specifically, we introduce a supervised multimodal variational autoencoder that can infer a meaningful latent space, where latent representations are placed along a disease trajectory. A score is computed by orthogonal projections onto this path. We evaluate our framework with multiple synthetic datasets and with a real dataset containing 14 patients, 40 presymptomatic genetic mutation carriers and 37 controls from the PREV-DEMALS study. There is no ground truth for the DPS in real-world scenarios, therefore we use the area under the ROC curve (AUC) as a proxy metric. Results with the synthetic datasets support this choice, since the higher the AUC, the more accurate the predicted simulated DPS. Experiments with the real dataset demonstrate better performance in comparison with stateof-the-art approaches. The proposed framework thus leverages cross-sectional multimodal datasets with small sample sizes to objectively measure disease progression, with potential application in clinical trials.

The determination of the number of mixture components (the order) of a finite mixture model has been an enduring problem in statistical inference. We prove that the closed testing principle leads to a sequential testing procedure (STP) that allows for confidence statements to be made regarding the order of a finite mixture model. We construct finite sample tests, via data splitting and data swapping, for use in the STP, and we prove that such tests are consistent against fixed alternatives. Simulation studies and real data examples are used to demonstrate the performance of the finite sample tests-based STP, yielding practical recommendations of their use as confidence estimators in combination with point estimates such as the Akaike information or Bayesian information criteria. In addition, we demonstrate that a modification of the STP yields a method that consistently selects the order of a finite mixture model, in the asymptotic sense. Our STP is not only applicable for order selection of finite mixture models, but is also useful for making confidence statements regarding any sequence of nested models.

Convolutional neural networks (CNN) have demonstrated their ability to segment 2D cardiac ultrasound images. However, despite recent successes according to which the intra-observer variability on end-diastole and end-systole images has been reached, CNNs still struggle to leverage temporal information to provide accurate and temporally consistent segmentation maps across the whole cycle. Such consistency is required to accurately describe the cardiac function, a necessary step in diagnosing many cardiovascular diseases. In this paper, we propose a framework to learn the 2D+time apical long-axis cardiac shape such that the segmented sequences can benefit from temporal and anatomical consistency constraints. Our method is a post-processing that takes as input segmented echocardiographic sequences produced by any state-of-the-art method and processes it in two steps to (i) identify spatio-temporal inconsistencies according to the overall dynamics of the cardiac sequence and (ii) correct the inconsistencies. The identification and correction of cardiac inconsistencies relies on a constrained autoencoder trained to learn a physiologically interpretable embedding of cardiac shapes, where we can both detect and fix anomalies. We tested our framework on 98 full-cycle sequences from the CAMUS dataset, which are available alongside this paper. Our temporal regularization method not only improves the accuracy of the segmentation across the whole sequences, but also enforces temporal and anatomical consistency., 12 pages, accepted for publication in IEEE TMI

International audience; Online atlasing, i.e., incrementing an atlas with new images as they are acquired, is key when performing studies on very large, or still being gathered, databases. Regular approaches to atlasing however do not focus on this aspect and impose a complete reconstruction of the atlas when adding images. We propose instead a diffeomorphic online atlasing method that allows gradual updates to an atlas. In this iterative centroid approach, we integrate new subjects in the atlas in an iterative manner, gradually moving the centroid of the images towards its final position. This leads to a computationally cheap approach since it only necessitates one additional registration per new subject added. We validate our approach on several experiments with three main goals: 1-to evaluate atlas image quality of the obtained atlases with sharpness and overlap measures, 2-to assess the deviation in terms of transformations with respect to a conventional atlasing method and 3-to compare its computational time with regular approaches of the literature. We demonstrate that the transformations divergence with respect to a stateof-the-art atlas construction method is small and reaches a plateau, that the two construction methods have the same ability to map subject homologous regions onto a common space and produce images of equivalent quality. The computational time of our approach is also drastically reduced for regular updates. Finally, we also present a direct extension of our method to update spatio-temporal atlases, especially useful for developmental studies.

International audience; Motion estimation in echocardiography plays an important role in the characterization of cardiac function, allowing the computation of myocardial deformation indices. However, there exist limitations in clinical practice, particularly with regard to the accuracy and robustness of measurements extracted from images. We therefore propose a novel deep learning solution for motion estimation in echocardiography. Our network corresponds to a modified version of PWC-Net which achieves high performance on ultrasound sequences. In parallel, we designed a novel simulation pipeline allowing the generation of a large amount of realistic B-mode sequences. These synthetic data, together with strategies during training and inference, were used to improve the performance of our deep learning solution, which achieved an average endpoint error of 0.07± 0.06 mm per frame and 1.20±0.67 mm between ED and ES on our simulated dataset. The performance of our method was further investigated on 30 patients from a publicly available clinical dataset acquired from a GE system. The method showed promise by achieving a mean absolute error of the global longitudinal strain of 2.5 ± 2.1% and a correlation of 0.77 compared to GLS derived from manual segmentation, much better than one of the most efficient methods in the state-of-the-art (namely the FFT-Xcorr block-matching method). We finally evaluated our method on an auxiliary dataset including 30 patients from another center and acquired with a different system. Comparable results were achieved, illustrating the ability of our method to maintain high performance regardless of the echocardiographic data processed.

In recent years, Empirical Bayesian (EB) inference has become an attractive approach for estimation in parametric models arising in a variety of real-life problems, especially in complex and high-dimensional scientific applications. However, compared to the relative abundance of available general methods for computing point estimators in the EB framework, the construction of confidence sets and hypothesis tests with good theoretical properties remains difficult and problem specific. Motivated by the universal inference framework of Wasserman et al. [2020], we propose a general and universal method, based on holdout likelihood ratios, and utilizing the hierarchical structure of the specified Bayesian model for constructing confidence sets and hypothesis tests that are finite sample valid. We illustrate our method through a range of numerical studies and real data applications, which demonstrate that the approach is able to generate useful and meaningful inferential statements in the relevant contexts.

The aim of this work was to compare the classification of cardiac MR-images of AL versus ATTR amyloidosis by neural networks and by experienced human readers. Cine-MR images and late gadolinium enhancement (LGE) images of 120 patients were studied (70 AL and 50 TTR). A VGG16 convolutional neural network (CNN) was trained with a 5-fold cross validation process, taking care to strictly distribute images of a given patient in either the training group or the test group. The analysis was performed at the patient level by averaging the predictions obtained for each image. The classification accuracy obtained between AL and ATTR amyloidosis was 0.750 for cine-CNN, 0.611 for Gado-CNN and between 0.617 and 0.675 for human readers. The corresponding AUC of the ROC curve was 0.839 for cine-CNN, 0.679 for gado-CNN (p < 0.004 vs. cine) and 0.714 for the best human reader (p < 0.007 vs. cine). Logistic regression with cine-CNN and gado-CNN, as well as analysis focused on the specific orientation plane, did not change the overall results. We conclude that cine-CNN leads to significantly better discrimination between AL and ATTR amyloidosis as compared to gado-CNN or human readers, but with lower performance than reported in studies where visual diagnosis is easy, and is currently suboptimal for clinical practice.

Submitted to Magnetic Resonance in Medicine; Purpose: To evaluate the benefits of SPARKLING, a stochastic k-space sampling approach over the deterministic sampling scheme TPI (Twisted Projection Imaging) in the context of accelerated cerebral $^{23}$Na MRI and assessing its impact on tissue sodium concentration (TSC) estimation.Methods: Guided by simulation results, in vitro and in vitro UTE $^{23}$Na MRI datasets were acquired at a 4 mm isotropic resolution from healthy volunteers using both TPI or SPARKLING trajectories with different acceleration factors (AF) on a 7T MR scanner equipped with a 32 channels head coil. Following reconstruction using NUFFT or Proximal Optimized Gradient Method (POGM) algorithms with or without regularization, respectively, the resulting sodium images were compared in terms of effective resolution (FWHM of the PSF), SNR and overall quality. with an in vitro assement of the accuracy of the TSC was performed via external referencing using a 4-point calibration approach. Results: In vivo cerebral sodium images acquired using SPARKLING with an acceleration factor of 32 (TA= 5 min 38 s) are similar to those obtained using TPI at AF=8 (TA= 22 min 34 s) with minimal impact on the accuracy of our TSC quantification. Conclusion: In conditions compatible with clinical examination, undersampled SPARKLING $^{23}$Na MRI can outperform the conventional TPI k-space sampling scheme allowing for shorter acquisition times.

Introduction: Primary Failure of Eruption (PFE) is a rare condition affecting posterior teeth eruption resulting in a posterior open bite malocclusion. Differential diagnosis like ankylosis or mechanical eruption failure should be considered. For non-syndromic forms, mutations in PTH1R, and recently in KMT2C genes are the known etiologies. The aim of this work was to describe the variability of clinical presentations of PFE associated with pathogenic variants of PTHR1.Material and methods: Diagnosis of non-syndromic PFE has been suggested for three members of a single family. Clinical and radiological features were collected, and genetic analyses were performed.Results: The clinical phenotype (type and number of involved teeth, depth of bone inclusions, functional consequences) is variable within the family. Severe tooth resorptions were detected. A heterozygous substitution in PTH1R (NM_000316.3): c.899T > C was identified as a class 4 likely pathogenic variant. The multidisciplinary management is described involving oral biology, pediatric dentistry, orthodontics, oral surgery, and prosthodontics.Conclusion: In this study, we report a new PTH1R variant involved in a familial form of PFE with variable expressivity. Therapeutic care is complex and difficult to systematize, hence the lack of evidence-based recommendations and clinical guidelines.

International audience; Unsupervised anomaly detection using deep learning models is a popular computer-aided diagnosis approach because it does not need annotated data and is not restricted to the diagnosis of a disease seen during training. Such approach consists in first learning the distribution of anomaly free images. Images presenting anomalies are then detected as outliers of this distribution. These approaches have been widely applied in neuroimaging to detect sharp and localized anomalies such as tumors or white matter hyper-intensities from structural MRI. In this work, we aim to detect anomalies from FDG PET images of patients with Alzheimer's disease. In this context, the anomalies can be subtle and difficult to delineate, making the task more difficult and meaning that no ground truth exists to evaluate the approaches. We thus propose a framework to evaluate unsupervised anomaly detection approaches that consists in simulating realistic anomalies from images of healthy subjects. We demonstrate the use of this framework by evaluating an approach based on a 3D variational autoencoder.

Automatic surgical instrument segmentation of endoscopic images is a crucial building block of many computer-assistance applications for minimally invasive surgery. So far, state-of-the-art approaches completely rely on the availability of a ground-truth supervision signal, obtained via manual annotation, thus expensive to collect at large scale. In this paper, we present FUN-SIS, a Fully-UNsupervised approach for binary Surgical Instrument Segmentation. FUN-SIS trains a per-frame segmentation model on completely unlabelled endoscopic videos, by solely relying on implicit motion information and instrument shape-priors. We define shape-priors as realistic segmentation masks of the instruments, not necessarily coming from the same dataset/domain as the videos. The shape-priors can be collected in various and convenient ways, such as recycling existing annotations from other datasets. We leverage them as part of a novel generative-adversarial approach, allowing to perform unsupervised instrument segmentation of optical-flow images during training. We then use the obtained instrument masks as pseudo-labels in order to train a per-frame segmentation model; to this aim, we develop a learning-from-noisy-labels architecture, designed to extract a clean supervision signal from these pseudo-labels, leveraging their peculiar noise properties. We validate the proposed contributions on three surgical datasets, including the MICCAI 2017 EndoVis Robotic Instrument Segmentation Challenge dataset. The obtained fully-unsupervised results for surgical instrument segmentation are almost on par with the ones of fully-supervised state-of-the-art approaches. This suggests the tremendous potential of the proposed method to leverage the great amount of unlabelled data produced in the context of minimally invasive surgery.

Most cerebrovascular diseases (including strokes and aneurysms) are treated endovascularly with cathetersthat are navigated from the groin through the vessels to the brain. Many patients have complex anatomy of the aortic arch and supra-aortic vessels, which can make it difficult to select the best catheters for navigation, resulting in longer procedures and more complications or failures. To this end, we propose a framework dedicated to the analysis of the aortic arch and supra-aortic trunks. This framework can automatically compute anatomical and geometrical features from meshes segmented beforehand via CNN-based pipeline. These features such as arch type, tortuosity and angulations describe the navigational difficulties encountered during catheterization. Quantitative and qualitative validation was performed by experienced neuroradiologists, leading to reliable vessel characterization.

International audience; Deep learning methods have become very popular for the processing of natural images, and were then successfully adapted to the neuroimaging field. As these methods are non-transparent, interpretability methods are needed to validate them and ensure their reliability. Indeed, it has been shown that deep learning models may obtain high performance even when using irrelevant features, by exploiting biases in the training set. Such undesirable situations can potentially be detected by using interpretability methods. Recently, many methods have been proposed to interpret neural networks. However, this domain is not mature yet. Machine learning users face two major issues when aiming to interpret their models: which method to choose, and how to assess its reliability? Here, we aim at providing answers to these questions by presenting the most common interpretability methods and metrics developed to assess their reliability, as well as their applications and benchmarks in the neuroimaging context. Note that this is not an exhaustive survey: we aimed to focus on the studies which we found to be the most representative and relevant.

Context-aware decision support in the operating room can foster surgical safety and efficiency by leveraging real-time feedback from surgical workflow analysis. Most existing works recognize surgical activities at a coarse-grained level, such as phases, steps or events, leaving out fine-grained interaction details about the surgical activity; yet those are needed for more helpful AI assistance in the operating room. Recognizing surgical actions as triplets of combination delivers comprehensive details about the activities taking place in surgical videos. This paper presents CholecTriplet2021: an endoscopic vision challenge organized at MICCAI 2021 for the recognition of surgical action triplets in laparoscopic videos. The challenge granted private access to the large-scale CholecT50 dataset, which is annotated with action triplet information. In this paper, we present the challenge setup and assessment of the state-of-the-art deep learning methods proposed by the participants during the challenge. A total of 4 baseline methods from the challenge organizers and 19 new deep learning algorithms by competing teams are presented to recognize surgical action triplets directly from surgical videos, achieving mean average precision (mAP) ranging from 4.2% to 38.1%. This study also analyzes the significance of the results obtained by the presented approaches, performs a thorough methodological comparison between them, in-depth result analysis, and proposes a novel ensemble method for enhanced recognition. Our analysis shows that surgical workflow analysis is not yet solved, and also highlights interesting directions for future research on fine-grained surgical activity recognition which is of utmost importance for the development of AI in surgery., Comment: CholecTriplet2021 challenge report. Paper accepted at Elsevier journal of Medical Image Analysis. 22 pages, 8 figures, 11 tables. Challenge website: https://cholectriplet2021.grand-challenge.org

Background and objective: 3D rotational angiography (3DRA) provides high quality images of the cerebral arteriovenous malformation (AVM) nidus that can be reconstructed in 3D. However, these reconstructions are limited to only 3D visualization without possible interactive exploration of geometric characteristics of cerebral structures. Refined understanding of the AVM angioarchitecture prior to treatment is mandatory and vascular segmentation is an important preliminary step that allow physicians analyze the complex vascular networks and can help guide microcatheters navigation and embolization of AVM.Methods: A deep learning method was developed for the segmentation of 3DRA images of AVM patients. The method uses a fully convolutional neural network with a U-Net-like architecture and a DenseNet backbone. A compound loss function, combining Cross Entropy and Focal Tversky, is employed for robust segmentation. Binary masks automatically generated from region-growing segmentation have been used to train and validate our model.Results: The developed network was able to achieve the segmentation of the vessels and the malformationand significantly outperformed the region-growing algorithm. Our experiments were performed on 9 AVMpatients. The trained network achieved a Dice Similarity Coefficient (DSC) of 80.43%, surpassing other U-Net like architectures and the region-growing algorithm on the manually approved test set by physicians.Conclusions: This work demonstrates the potential of a learning-based segmentation method for characterizing very complex and tiny vascular structures even when the training phase is performed with the results of an automatic or a semi-automatic method. The proposed method can contribute to the planning and guidance of endovascular procedures.

Aims Outcomes in pulmonary hypertension (PH) are related to right ventricular (RV) function and remodelling. We hypothesized that changes in RV function and especially area strain (AS) could provide incremental prognostic information compared to the use of baseline data only. We therefore aimed to assess RV function changes between baseline and 6-month follow-up and evaluate their prognostic value for PH patients using 3D echocardiography. Methods and results Ninety-five PH patients underwent a prospective longitudinal study including ESC/ERS guidelines prognostic assessment and 3D RV echocardiographic imaging at baseline and 6-month follow-up. Semi-automatic software tracked the RV along the cycle, and its output was post-processed to extract 3D deformation patterns. Over a median follow-up of 24.8 (22.1–25.7) months, 21 patients died from PH or were transplanted. Improvements in RV global AS were associated with stable or improving clinical condition as well as survival free from transplant (P < 0.001). The 3D deformation patterns confirmed that the most significant regional changes occurred within the septum. RV global AS change over 6-month by +3.5% identifies patients with a 3.7-fold increased risk of death or transplant. On multivariate COX analysis, changes in WHO class, BNP, and RV global AS were independent predictors of outcomes. Besides, the combination of these three parameters was of special interest to identify high-risk patients [HR 11.5 (1.55–86.06)]. Conclusion Changes in RV function and especially changes in 3D RV AS are of prognostic importance. Our study underlines that assessing such changes from baseline to follow-up is of additional prognostic value for PH patients. Clinical Trial Registration http://clinicaltrials.gov/ct2/show/NCT02799979

Containing the medical data of millions of patients, clinical data warehouses (CDWs) represent a great opportunity to develop computational tools. Magnetic resonance images (MRIs) are particularly sensitive to patient movements during image acquisition, which will result in artefacts (blurring, ghosting and ringing) in the reconstructed image. As a result, a significant number of MRIs in CDWs are corrupted by these artefacts and may be unusable. Since their manual detection is impossible due to the large number of scans, it is necessary to develop tools to automatically exclude (or at least identify) images with motion in order to fully exploit CDWs. In this paper, we propose a novel transfer learning method for the automatic detection of motion in 3D T1-weighted brain MRI. The method consists of two steps: a pre-training on research data using synthetic motion, followed by a fine-tuning step to generalise our pre-trained model to clinical data, relying on the labelling of 4045 images. The objectives were both (1) to be able to exclude images with severe motion, (2) to detect mild motion artefacts. Our approach achieved excellent accuracy for the first objective with a balanced accuracy nearly similar to that of the annotators (balanced accuracy>80 %). However, for the second objective, the performance was weaker and substantially lower than that of human raters. Overall, our framework will be useful to take advantage of CDWs in medical imaging and highlight the importance of a clinical validation of models trained on research data.

Quasi-static ultrasound elastography is a medical imaging technique dedicated to the mechanical characterization of biological tissues. This technique uses, for diagnosis, criteria mainly derived from the axial strain, which is the most accurately estimated component. However, imaging the strain tensor would be of interest, not only for a better evaluation of the tissue response to compression, but also for the implementation of reconstruction methods to access to the mechanical properties of the examined tissues.In this PhD work, two research axes were developed to analyze the elastic properties of biological tissues. The first axis focused on the development of a regularization method in order to smooth the displacement fields and to obtain less noisy, and thus usable, strain images. The results on numerical simulations, phantoms and breast tissues showed the strong contribution of this method, when facing complex and noisy data, especially for the lateral components of displacement and strain. Moreover, the proposed approach allowed to obtain shear strain images and rotation elastograms on which shear/rotation patterns were clearly detectable around the inclusions/lesions. The second research axis focused on the development of a method using the virtual work principle to reconstruct the spatial distribution of the Young's modulus within linear, elastic and isotropic media. This method requires the knowledge of the displacement/strain fields and the applied force. To overcome the lack of 3D ultrasound data, the reconstruction of the elastic modulus has been studied under the assumptions of plane stresses and plane strains. The results on numerical simulations and phantoms showed an accurate reconstruction of the Young's modulus mapping, in terms of region localization, shape and size. The modulus values are correctly estimated when the assumptions introduced in the method are respected. On the other hand, when the assumptions are no longer satisfied, a decrease in elastic contrast is observed. Nevertheless, the regions with distinct elastic properties in a medium remain clearly visible, both in simulations and in phantom data. Finally, the results on breast tissues indicate a stiffer region in agreement with the position of the lesion observable on the ultrasound imaging. These different results are promising and particularly interesting for a more complete analysis of biological tissues in quasi-static ultrasound elastography., L'élastographie ultrasonore quasi-statique est une technique d'imagerie médicale dédiée à la caractérisation mécanique des tissus biologiques. Cette technique utilise pour le diagnostic des critères issus principalement des images de la déformation axiale, qui est la composante estimée avec le plus de précision. Néanmoins, une imagerie complète du tenseur de déformation serait intéressante, non seulement pour une meilleure évaluation de la réponse du tissu à la compression, mais aussi et surtout pour la mise en œuvre de méthodes de reconstruction permettant d'accéder aux propriétés mécaniques des tissus imagés.Dans ces travaux de thèse, deux axes de recherche ont été développés afin de permettre une analyse des propriétés élastiques des tissus biologiques. Le premier axe a porté sur le développement d'une méthode de régularisation afin de lisser les champs de déplacement et d'obtenir des images de déformation moins bruitées, et donc exploitables. Les résultats sur des simulations numériques, des fantômes et des tissus mammaires in vivo, ont montré la forte contribution de cette méthode face à des données complexes et bruitées, notamment pour les composantes latérales du déplacement et de la déformation. De plus, l'approche proposée a permis d'obtenir des images de déformation en cisaillement et des élastogrammes de rotation sur lesquels des motifs de cisaillement/rotation étaient clairement détectables autour des inclusions/lésions. Le deuxième axe de recherche a été centré sur le développement d'une méthode utilisant le principe des travaux virtuels pour reconstruire la distribution spatiale du module d'Young au sein de milieux élastiques, linéaires et isotropes. Cette méthode nécessite de connaitre au préalable les champs de déplacement/déformation ainsi que la force appliquée. Pour pallier le manque de données 3D en échographie, la reconstruction du module élastique a été étudiée sous les hypothèses des contraintes et déformations planes. Les résultats sur simulations numériques et fantômes ont montré une reconstruction précise de la cartographie du module d'Young, en termes de localisation des régions, de leur forme et de leur taille. Les valeurs de module sont correctement estimées lorsque les hypothèses introduites dans la méthode sont respectées. En revanche, lorsque les hypothèses ne sont plus satisfaites, une diminution du contraste élastique est observée. Néanmoins les régions aux propriétés élastiques distinctes dans un milieu restent clairement visibles, que ce soit en simulations ou pour les données de fantômes. Enfin, les résultats sur tissus mammaires indiquent une région plus rigide en accord avec la position de la lésion observable à l'échographie. Ces différents résultats sont prometteurs et particulièrement intéressants pour une analyse plus complète des tissus biologiques en élastographie ultrasonore quasi-statique.

Neuroimaging offers an unmatched description of the brain’s structure and physiology, but the information it provides is not easy to extract and interpret. A popular way to extract meaningful information from brain images is to use computational methods based on machine learning and deep learning to predict the current or future diagnosis of a patient. A large number of these approaches have been dedicated to the computer-aided diagnosis of dementia, and more specifically of Alzheimer's disease. However, only a few are translated to the clinic. This can be explained by different factors such as the lack of rigorous validation of these approaches leading to over-optimistic performance and their lack of reproducibility, but also the limited interpretability of these methods and their limited generalisability when moving from highly controlled research data to routine clinical data. This manuscript describes how we tried to address these limitations.We have proposed reproducible frameworks for the evaluation of Alzheimer's disease classification methods and developed two open-source software platforms for clinical neuroimaging studies (Clinica) and neuroimaging processing with deep learning (ClinicaDL). We have implemented and assessed the robustness of a visualisation method aiming to interpret convolutional neural networks and used it to study the stability of the network training. We concluded that, currently, combining a convolutional neural networks classifier with an interpretability method may not constitute a robust tool for individual computer-aided diagnosis. As an alternative, we have proposed an approach that detects anomalies in the brain by generating what would be the healthy version of a patient's image and comparing this healthy version with the real image. Finally, we have studied the performance of machine and deep learning algorithms for the computer-aided diagnosis of dementia from images acquired in clinical routine., La neuro-imagerie offre une description inégalée de la structure et de la physiologie du cerveau, mais les informations qu'elle fournit ne sont pas faciles à extraire et à interpréter. Une façon populaire d'extraire des informations pertinentes d'images cérébrales consiste à utiliser des méthodes basées sur l'apprentissage statistique et l'apprentissage profond pour prédire le diagnostic actuel ou futur d'un patient. Un grand nombre de ces approches ont été dédiées au diagnostic assisté par ordinateur de la démence, et plus spécifiquement de la maladie d'Alzheimer. Cependant, seules quelques-unes sont transposées en clinique. Cela peut s'expliquer par différents facteurs tels que l'absence de validation rigoureuse de ces approches conduisant à des performances trop optimistes et à leur manque de reproductibilité, mais aussi l'interprétabilité limitée de ces méthodes et leur généralisation limitée lors du passage de données de recherche hautement contrôlées à des données cliniques de routine. Ce manuscrit décrit comment nous avons tenté de remédier à ces limites.Nous avons proposé des cadres reproductibles pour l'évaluation des méthodes de classification de la maladie d'Alzheimer et développé deux plateformes logicielles open-source pour les études de neuroimagerie clinique (Clinica) et le traitement de la neuroimagerie par apprentissage profond (ClinicaDL). Nous avons implémenté et évalué la robustesse d'une méthode de visualisation visant à interpréter les réseaux neuronaux convolutifs et l'avons utilisée pour étudier la stabilité de l'entraînement du réseau. Nous avons conclu qu'actuellement, la combinaison de réseaux neuronaux convolutifs avec une méthode d'interprétabilité peut ne pas constituer un outil robuste pour le diagnostic individuel assisté par ordinateur. De façon alternative, nous avons proposé une approche qui détecte les anomalies dans le cerveau en générant ce qui serait la version saine de l'image d'un patient et en comparant cette version saine avec l'image réelle. Enfin, nous avons étudié les performances des algorithmes d'apprentissage statistique et profond pour le diagnostic assisté par ordinateur de la démence à partir d'images acquises en routine clinique.

White matter fibers transfer information between brain regions with delays that are observable with magnetoencephalography and electroencephalography (M/EEG) due to their millisecond temporal resolution. We can represent the brain as a graph where nodes are the cortical sources or areas and edges are the physical connections between them: either local (between adjacent vertices on the cortical mesh) or non-local (long-range white matter fibers). Long-range anatomical connections can be obtained with diffusion MRI (dMRI) tractography which yields a set of streamlines representing white matter fiber bundles. Given the streamlines’ lengths and the information conduction speed, transmission delays can be estimated for each connection. dMRI can thus give an insight into interaction delays of the macroscopicbrain network.Localizing and recovering electrical activity of the brain from M/EEG measurements is known as the M/EEG inverse problem. Generally, there are more unknowns (brain sources) than the number of sensors, so the solution is non-unique and the problem ill-posed. To obtain a unique solution, prior constraints on the characteristics of source distributions are needed. Traditional linear inverse methods deploy different constraints which can favour solutions with minimum norm, impose smoothness constraints in space and/or time along the cortical surface, etc. Yet, structural connectivity is rarely considered and transmission delays almost always neglected.The first contribution of this thesis consists of a multimodal preprocessing pipeline used to integrate structural MRI, dMRI and MEG data into a same framework, and of a simulation procedure of source-level brain activity that was used as a synthetic dataset to validate the proposed reconstruction approaches.In the second contribution, we proposed a new framework to solve the M/EEG inverse problem called Connectivity-Informed M/EEG Inverse Problem (CIMIP), where prior transmission delays supported by dMRI were included to enforce temporal smoothness between time courses of connected sources. This was done by incorporating a Laplacian operator into the regularization, that operates on a time-dependent connectivity graph. Nonetheless, some limitations of the CIMIP approach arised, mainly due to the nature of the Laplacian, which acts on the whole graph, favours smooth solutions across all connections, for all delays, and it is agnostic to directionality.In this thesis, we aimed to investigate patterns of brain activity during visuomotor tasks, during which only a few regions typically get significantly activated, as shown by previous studies. This led us to our third contribution, an extension of the CIMIP approach that addresses the aforementioned limitations, named CIMIP_OML (“Optimal Masked Laplacian”). We restricted the full source space network (the whole cortical mesh) to a network of regions of interest and tried to find how the information is transferred between its nodes. To describe the interactions between nodes in a directed graph, we used the concept of network motifs. We proposed an algorithm that (1) searches for an optimal network motif – an optimal pattern of interaction between different regions and (2) reconstructs source activity given the found motif. Promising results are shown for both simulated and real MEG data for a visuomotor task and compared with 3 different state-of-the-art reconstruction methods.To conclude, we tackled a difficult problem of exploiting delays supported by dMRI for the reconstruction of brain activity, while also considering the directionality in the information transfer, and provided new insights into the complex patterns of brain activity., Les fibres de la matière blanche permettent le transfert d’information dans le cerveau avec des délais observables en Magnétoencéphalographie et Électroencéphalographie (M/EEG) grâce à leur haute résolution temporelle. Le cerveau peut être représenté comme un graphe où les nœuds sont les régions corticales et les liens sont les connexions physiques entre celles-ci: soit locales (entre sommets adjacents sur le maillage cortical), soit non locales (fibres de la matière blanche). Les connexions non-locales peuvent être reconstruites avec la tractographie de l’IRM de diffusion (IRMd) qui génère un ensemble de courbes («streamlines») représentant des fibres de la matière blanche. Sachant les longueurs des fibres et la vitesse de conduction de l’information, les délais de transmission peuvent être estimés. L’IRMd peut donc donner un aperçu des délais d’interaction du réseau cérébral macroscopique.La localisation et la reconstruction de l’activité électrique cérébrale à partir des mesures M/EEG est un problème inverse. En général, il y a plus d’inconnues (sources cérébrales) que de capteurs. La solution n’est donc pas unique et le problème est dit mal posé. Pour obtenir une solution unique, des hypothèses sur les caractéristiques des distributions de sources sont requises. Les méthodes inverses linéaires traditionnelles utilisent différentes hypothèses qui peuvent favoriser des solutions de norme minimale, imposer des contraintes de lissage dans l’espace et/ou dans le temps, etc. Pourtant, la connectivité structurelle est rarement prise en compte et les délais de transmission sont presque toujours négligés.La première contribution de cette thèse est un pipeline de prétraitement multimodal utilisé pour l’intégration des données d’IRM, IRMd et MEG dans un même cadre, et d’une méthode de simulation de l’activité corticale qui a été utilisée comme jeu de données synthétiques pour valider les approches de reconstruction proposées. Nous proposons également une nouvelle approche pour résoudre le problème inverse M/EEG appelée «Problème Inverse M/EEG Informé par la Connectivité» (CIMIP pour Connectivity-Informed M/EEG Inverse Problem), où des délais de transmission provenant de l’IRMd sont inclus pour renforcer le lissage temporel entre les décours des sources connectées. Pour cela, un opérateur Laplacien, basé sur un graphe de connectivité en fonction du temps, a été intégré dans la régularisation. Cependant, certaines limites de l’approche CIMIP sont apparues en raison de la nature du Laplacien qui agit sur le graphe entier et favorise les solutions lisses sur toutes les connexions, pour tous les délais, et indépendamment de la directionnalité.Lors de tâches visuo-motrices, seules quelques régions sont généralement activées significativement. Notre troisième contribution est une extension de CIMIP pour ce type de tâches qui répond aux limitations susmentionnées, nommée CIMIP_OML («Optimal Masked Laplacian») ou Laplacien Masqué Optimal. Nous essayons de trouver comment l’information est transférée entre les nœuds d’un sous-réseau de régions d’intérêt du réseau complet de l’espace des sources. Pour décrire les interactions entre nœuds dans un graphe orienté, nous utilisons le concept de motifs de réseau. Nous proposons un algorithme qui 1) cherche un motif de réseau optimal- un modèle optimal d’interaction entre régions et 2) reconstruit l’activité corticale avec le motif trouvé. Des résultats prometteurs sont présentés pour des données MEG simulées et réelles (tâche visuo-motrice) et comparés avec 3 méthodes de l’état de l’art. Pour conclure, nous avons abordé un problème difficile d’exploitation des délais de l’IRMd lors l’estimation de l’activité corticale en tenant compte de la directionalité du transfert d’information, fournissant ainsi de nouvelles perspectives sur les patterns complexes de l’activité cérébrale.

In medical imaging, segmentation using deep learning enables an automatic generation of anatomical models that are crucial for morphological evaluation. However, the scarcity of pediatric imaging resources may result in reduced accuracy and generalization performance of segmentation networks. To mitigate these issues, our first approach consists in a novel optimization scheme leveraging shape priors to enforce globally consistent predictions and an adversarial network to encourage precise delineations. In our second strategy, we design a novel multi-task, multi-domain network optimized over multi-anatomy imaging datasets. To improve generalizability, we disentangle the domains representations using a contrastive regularization and extend the shape priors to multi-anatomy learning. Our contributions are evaluated for the bone segmentation of three anatomical joints (ankle, knee, shoulder). The proposed methods performed either better or at par with state-of-the-art models. These results bring new perspectives towards a collaborative utilization of pediatric imaging resources and better management of musculoskeletal disorders., En imagerie médicale, la segmentation basée sur l’apprentissage profond permet de générer automatiquement des modèles anatomiques qui sont cruciaux pour l’évaluation morphologique. Cependant, la rareté des ressources d’imagerie pédiatrique peut entraîner une diminution de la précision et des performances de généralisation des réseaux de segmentation. Pour atténuer ces problèmes, notre première approche consiste en un nouveau schéma d’optimisation exploitant des à priori de formes visant à imposer des prédictions globalement cohérentes et un réseau antagoniste qui encourage des délimitations plus précises. Dans notre deuxième stratégie, nous concevons un nouveau réseau multi-tâche et multi-domaine optimisé sur des ensembles de données d’imagerie multi-anatomie. Pour améliorer la généralisation, nous démêlons les représentations des domaines en utilisant une régularisation contrastive et nous étendons les à priori de formes à l’apprentissage multi-anatomie. Nos contributions sont évaluées pour la segmentation osseuse de trois articulations (cheville, épaule, genou). Les méthodes proposées ont obtenu des résultats supérieurs ou égaux à ceux des modèles de l’état de l’art. Ces résultats ouvrent de nouvelles perspectives pour une utilisation collaborative des ressources d’imagerie pédiatrique et une meilleure gestion des troubles musculo-squelettiques.

International audience

Introduction: Magnetic resonance imaging (MRI) plays a central role at the time of diagnosis and throughout the follow-up of patients with glioblastoma (GBM). Perfusion-weighted imaging after injection of gadolinium makes it possible to assess tumor neoangiogenesis by calculating perfusion parameters, such as relative cerebral blood volume (rCBV), correlated with tumor grade and patient prognosis. The use of this sequence is recommended during the follow-up of GBMs, as treatments may be the cause of phenomena mimicking tumor progression, such as pseudoprogression (PsP). Its diagnosis, both in imaging and in histology, remains a challenge in neuro-oncology. The classical perfusion-MRI analysis method, by selection of regions of interest (ROI), has a high inter-observer variability and is responsible for lesion sampling, potentially detrimental for heterogeneous lesions such as GBM or PsP. However, no alternative method is offered in clinical practice. From a volume analysis of perfusion-MRI, we developed a perfusion parameter evaluating the hypervascularized tumor fraction, then assessed its reproducibility and clinical interest during the follow-up of patients with GBM. Methods: Our parameter was calculated after manual segmentation of the contrast-enhanced lesion on the post-contrast 3DT1-weighted images, then extraction of the perfusion data from all the pixels included in the segmented volume and finally calculation of the proportion of these pixels possessing a value of rCBV above a predefined threshold, in this case 2. The reproducibility of this fractional parameter, called %rCBV>2, was evaluated by the intra-class correlation coefficient (ICC) and the coefficient of variation (CV), after 3 operators performed its measurement on MRIs of 27 newly diagnosed GBMs. We then analyzed its performance for the diagnosis of PsP, on a retrospective cohort of patients with progressive lesion on the 1st post-treatment follow-up MRI. Finally, from a multicenter and prospective cohort, we measured this parameter on the 1st follow-up MRI of all the included patients and studied its correlation with early progression and patient survival. Results: The measurement of %rCBV>2 had excellent inter- and intra-observer reproducibility (ICC=0.94 and 0.91, respectively; CV15%). In our 1st cohort of 25 patients, %rCBV>2 was correlated with the development of PsP, whether calculated on the MRI performed at diagnosis (PsP: 57.5% vs. progression: 71.3%; p=0.033) or on the 1st follow-up MRI (PsP: 22.1% vs. progression: 51.8%; p=0.0006; AUC=0.909). In the 2nd cohort of 46 patients, no correlation was demonstrated between %rCBV>2 and early progression or survival. Conclusion: We have developed a volume analysis of perfusion-MRI to assess the hypervascularized fraction of a lesion using the fractional parameter: %rCBV>2. Its reproducibility was excellent, superior to the classic ROI method. While its results were promising for establishing the diagnosis of PsP on early follow-up MRI of a retrospective cohort, we were unable to show any correlation between the measurement of this parameter and tumor progression or patient survival on a multicenter prospective cohort. Still, this fractional parameter appears very reproducible and potentially more representative than the classic rCBV to assess the perfusion of heterogeneous lesions such as GBM. However, its use in clinical practice as an MRI biomarker still requires validation steps on larger, multicenter and prospective cohorts., Introduction : L'imagerie par résonance magnétique (IRM) est indispensable au moment du diagnostic et tout au long du suivi des patients atteints de glioblastome (GBM). L'IRM de perfusion après injection de gadolinium permet d'apprécier la néoangiogenèse tumorale par le calcul de paramètres perfusionnels, comme le volume sanguin cérébral relatif (rCBV), corrélé au grade tumoral ou encore au pronostic du patient. L'utilisation de cette séquence est recommandée au cours du suivi des GBM, les traitements pouvant être à l'origine de phénomènes mimant une progression tumorale, comme la pseudoprogression (PsP). Son diagnostic, tant en imagerie qu'en histologie, reste un enjeu en neuro-oncologie. L'analyse classique de la séquence de perfusion, par sélection de régions d'intérêt (ROI), possède une importante variabilité inter-observateur et est à l'origine d'un échantillonnage de la lésion, potentiellement préjudiciable pour des lésions hétérogènes comme les GBM ou la PsP. Cependant, aucune méthode alternative n'est proposée en pratique clinique. A partir d'une analyse volumique de l'IRM de la perfusion, nous avons développé un paramètre perfusionnel évaluant la fraction tumorale hypervascularisée, puis avons évalué sa reproductibilité et son intérêt clinique au cours du suivi de patients atteints de GBM. Méthodes : Notre paramètre était calculé après segmentation manuelle du volume de la prise de contraste sur la séquence 3DT1 injectée, puis extraction des données de perfusion de l'ensemble des pixels inclus dans le volume segmenté et enfin calcul de la proportion de ces pixels possédant une valeur de rCBV au-dessus d'un seuil prédéfini, en l'occurrence 2. La reproductibilité de ce paramètre fractionnel, appelé %rCBV>2, a été évaluée par le coefficient de corrélation intra-classe (ICC) et le coefficient de variation (CV) après que 3 opérateurs ont eu réalisé sa mesure sur les IRM de 27 GBM nouvellement diagnostiqués. Nous avons ensuite analysé ses performances pour le diagnostic de PsP, sur une cohorte rétrospective de patients atteints de GBM en progression sur la 1ère IRM de suivi post-thérapeutique. Enfin, à partir d'une cohorte prospective et multicentrique, nous avons mesuré ce paramètre sur la 1ère IRM de suivi de l'ensemble des patients et étudié sa corrélation avec l'évolutivité précoce ainsi que la survie des patients. Résultats : La mesure du %rCBV>2 possédait une excellente reproductibilité inter- et intra-observateur (ICC=0.94 et 0.91 respectivement; CV15%). Sur notre 1ère cohorte de 25 patients, le %rCBV>2 était corrélé au développement d'une PsP, qu'il soit calculé sur l'IRM réalisée au diagnostic (PsP: 57.5% vs progression : 71.3%; p=0.033) ou sur la 1ère IRM de suivi (PsP : 22.1% vs progression: 51.8%; p=0.0006; AUC=0.909). Sur la 2ème cohorte de 46 patients, aucune corrélation n'a été mise en évidence entre le %rCBV>2 et l'évolutivité précoce ou la survie. Conclusion : Nous avons développé une méthode d'analyse volumique de l'IRM de perfusion permettant d'évaluer la fraction hypervascularisée d'une lésion grâce au paramètre fractionnel %rCBV>2. Sa reproductibilité était excellente, supérieure à la méthode classique d'analyse par ROI. Si ses résultats étaient prometteurs pour établir le diagnostic de PsP sur l'IRM de suivi précoce d'une cohorte rétrospective, nous n'avons pas pu montrer de corrélation entre la mesure de ce paramètre et l'évolutivité tumorale ou la survie des patients sur une cohorte prospective multicentrique. Ce paramètre fractionnel apparait toutefois très reproductible et potentiellement plus représentatif que le classique rCBV pour apprécier la perfusion de lésions hétérogènes tels que les GBM. Son utilisation en pratique clinique en tant que biomarqueur en IRM nécessite cependant encore des étapes de validation sur des cohortes plus larges, multicentriques et prospectives.

International audience; We describe a new procedure for three-dimensional (3D) region-of-interest (ROI) reconstruction from transverselytruncated cone-beam projections acquired with a circular source trajectory. This method is an extension to 3D of the virtual fan-beam (VFB) method. It is based on a VFB formula that performs the backprojection in the acquisition geometry. Our simulation results show that the ROI reconstruction of the 3D Shepp-Logan phantom is very similar to the one obtained by the Feldkamp, Davis, Kress (FDK) algorithm without truncation. However the reconstruction of the Forbild head phantom shows artefacts which are absent from the FDK truncation-free reconstruction.

International audience; Preoperative path planning for Deep Brain Stimulation (DBS) is a multi-objective optimization problem consisting in searching the best compromise between multiple placement constraints. Its automation is usually addressed by turning the problem into mono objective thanks to an aggregative approach. However,despite its intuitiveness,this approach is known for its incapacity to find all optimal solutions. In this work,we introduce an approach based on multi objective dominance to DBS path planning. We compare it to a classical aggregative weighted sum of the multiple constraints and to a manual planning thanks to a retrospective study performed by a neurosurgeon on 14 DBS cases. The results show that the dominance-based method is preferred over manual planning,and covers a larger choice of relevant optimal entry points than the traditional weighted sum approach which discards interesting solutions that could be preferred by surgeons. © Springer International Publishing AG 2016.

International audience

International audience; The extraction of blood vessels has recently experienced a widespread interest in medical image analysis. Automatic vessel segmentation is highly desirable to guide clinicians in computer-assisted diagnosis, therapy or surgical planning. Despite a good ability to extract large anatomical structures, the capacity of U-Net inspired architectures to automatically delineate vascular systems remains a major issue, especially given the scarcity of existing datasets. In this paper, we present a novel approach that integrates into deep segmentation shape priors from a Semi-Overcomplete Convolutional Auto-Encoder (S-OCAE) embedding. Compared to standard Convolutional Auto-Encoders (CAE), it exploits an over-complete branch that projects data onto higher dimensions to better characterize tiny structures. Experiments on retinal and liver vessel extraction, respectively performed on publicly-available DRIVE and 3D-IRCADb datasets, highlight the effectiveness of our method compared to U-Net trained without and with shape priors from a traditional CAE.

International audience; Weaning a patient from mechanical ventilation is a critical task in Intensive Care Units, but it can be made safer by using Lung Ultrasound scoring. This scoring is currently done visually by specialists based on Lung Ultrasound artifacts among which are A-lines. Automating this scoring may help standardizing results and saving time for clinicians. In this paper, we propose a method to automatically detect A-lines from a manual delineation of the pleural line, and by using both the intensity profile of the LUS image and morphological operations. A score is then assigned to significant lines and represents the possibility of them being A-lines. The proposed method shows promising results in differentiating A-lines from other elements with an Area Under the Curve of 0.95; furthermore, using a threshold at 0.5 to detect A-lines leads to very good performances with an accuracy of 95% and a F0.5-score of 0.84.

International audience; The intima-media complex of the common carotid artery is considered the sentinel of a silent killer disease called atherosclerosis. Morphological biomarkers such as the intima-media thickness are already exploitable, but dynamic biomarkers, which reflect tissue deformation over the cardiac cycle, remain to be validated. Recent motion estimation methods seek to quantify compression, shear, and elongation coefficients, but their clinical applicability has not yet been well defined, and their actual accuracy is difficult to assess due to the absence of ground truth. This lack of reference also is the main limitation to explore fully supervised deep learning methods that have shown great potential in other applications. With this in mind, we propose a simulation pipeline to produce realistic in silico sequences, by combining a physics-based simulator with a post-processing based on a generative adversarial network.

Nuclear medicine is a medical specialty that studies the physiology of organs and the metabo-lism of various types of tumors. Nuclear medicine uses pharmaceuticals bound to a radioactive isotope. Molecular radiotherapy (MRT) is a specialty of nuclear medicine where the vector is directed to targets, usually tumors, and the action of ionizing radiation is aimed at destroying tumors. The follow-up and optimization of MRT requires the evaluation of the irradiation delivered to the patient (dosimetry). There is a lack of standardization in internal dosimetry. This thesis provides a standardized approach with descriptive clinical dosimetry workflows. A software named OpenDose3D, based in 3D-Slicer and implementing the proposed workflows was developed, validated and was made publicly available as an open source module. The module was used in clinical research within the MEDIRAD project., La médecine nucléaire est une spécialité médicale qui étudie la physiologie des organes et le métabolisme de divers types de tumeurs. La médecine nucléaire utilise des produits pharmaceutiques liés à un isotope radioactif. La radiothérapie interne vectorisée (RIV) est une spécialité de la médecine nucléaire où le vecteur est dirigé vers des cibles, généralement des tumeurs, et où l'action des rayonnements ionisants vise à détruire les tumeurs. Le suivi et l'optimisation de la RMT nécessitent l'évaluation de l'irradiation délivrée au patient (dosimétrie). Il y a un manque de standardisation en dosimétrie interne. Cette thèse propose une approche standardisée avec des flux de travail descriptifs pour la dosimétrie clinique. Un logiciel appelé OpenDose3D, basé sur 3D-Slicer en tant que module open source mettant en oeuvre les flux de travail proposés, est développé, validé et mis à la disposition du public. Le module a été utilisé en recherche clinique dans le projet MEDIRAD.

International audience; Purpose: We propose to learn a 3D keypoint descriptor which we use to match keypoints extracted from full-body CT scans. Our methods are inspired by 2D keypoint descriptor learning, which was shown to outperform hand-crafted descriptors. Adapting these to 3D images is challenging because of the lack of labelled training data and high memory requirements.Method: We generate semi-synthetic training data. For that, we first estimate the distribution of local affine inter-subject transformations using labelled anatomical landmarks on a small subset of the database. We then sample a large number of transformations and warp unlabelled CT scans, for which we can subsequently establish reliable keypoint correspondences using guided matching. These correspondences serve as training data for our descriptor, which we represent by a CNN and train using the triplet loss with online triplet mining.Results: We carry out experiments on a synthetic data reliability benchmark and a registration task involving 20 CT volumes with anatomical landmarks used for evaluation purposes. Our learned descriptor outperforms the 3D-SURF descriptor on both benchmarks while having a similar runtime.Conclusion: We propose a new method to generate semi-synthetic data and a new learned 3D keypoint descriptor. Experiments show improvement compared to a hand-crafted descriptor. This is promising as literature has shown that jointly learning a detector and a descriptor gives further performance boost.

Background : VIP (the Virtual Imaging Platform) is a web portal for medical imaging (MI) data analysis. By leveraging computational and storage resources from the EGI e-infrastructure, VIP provides MI researchers with end-user services to run MI applications on this large-scale computing infrastructure. Research Issue : Medical imaging is facing a reproducibility crisis: the increasing complexity of current data processing methods weakens our ability to produce the same results twice, by applying the same treatments to the same sets of inputs. Beyond the trivial influence of the data exploration process on any scientific result, there is mounting evidence that computing environments (e.g., library calls, OS kernels, hardware infrastructures) also play a significant role by adding numerical uncertainty. Relying on distributed computing resources, the VIP platform is highly concerned by potential versatilities in its digital outcomes. Project Outline : The ReproVIP project addresses this reproducibility issue at every level of data analysis, from the exploration process to the computing environment. It is structured around two complementary goals: (i) evaluate the uncertainty of digital outcomes after EGI-based distributed computing, and (ii) enhance the numerical reproducibility of scientific results obtained through the VIP platform.

International audience; Major depression is a leading cause of disability due to its trend to recurrence and treatment resistance. Currently, there are no biomarkers which could potentially identify patients with risk of treatment resistance. In this original paper, we propose a two-level shape analysis of the white matter bundles based on the Large Diffeomorphic Deformation Metric Mapping framework, to study treatment resistant depression. Fiber bundles are characterised via the deformation of their center line from a centroid shape. We developed two statistical analyses at a global and a local level to identify the most relevant bundles related to treatment resistant depression. Using a prospective longitudinal cohort including 63 patients. We applied this approach at baseline on 50 white matter fiber-tracts, to predict the clinical improvement at 6 months. Our results show a strong association between three bundles and the clinical improvement 6 months after. More precisely, the rightsided thalamo-occipital fascicle and optic radiations are the most robust followed by the splenium. The present study shows the interest in considering white matter shape in the context of depression, contributing to improve our understanding of neurobiological process of treatment resistance depression.

Video-assisted thoracoscopic surgery (VATS) is a minimally invasive surgical technique for the diagnosis and treatment of early-stage lung cancer. During VATS, large lung deformation occurs as a result of a change of patient position and a pneumothorax (lung deflation), which hinders the intraoperative localization of pulmonary nodules. Modeling lung deformation during VATS for surgical navigation is desirable, but the mechanisms causing such deformation are yet not well-understood. In this study, we estimate, quantify and analyze the lung deformation occurring after a change of patient position during VATS. We used deformable image registration to estimate the lung deformation between a preoperative CT (in supine position) and an intraoperative CBCT (in lateral decubitus position) of six VATS clinical cases. We accounted for sliding motion between lobes and against the thoracic wall and obtained consistently low average target registration errors (under 1 mm). We observed large lung displacement (up to 40 mm); considerable sliding motion between lobes and against the thoracic wall (up to 30 mm); and localized volume changes indicating deformation. These findings demonstrate the complexity of the change of patient position phenomenon, which should necessarily be taken into account to model lung deformation for intraoperative guidance during VATS., Comment: Accepted for publication in IEEE Transactions of Biomedical Engineering

International audience; Segmentation of brain tumor images, to refine the detection and understanding of abnormal masses in the brain, is an important research topic in medical imaging. This paper proposes a new segmentation method, consisting of three main steps, to detect brain lesions using magnetic resonance imaging (MRI). In the first step, the parts of the image delineating the skull bone are removed, to exclude insignificant data. In the second step, which is the main contribution of this study, the particle swarm optimization (PSO) technique is applied, to detect the block that contains the brain lesions. The fitness function, used to determine the best block among all candidate blocks, is based on a two-way fixed-effects analysis of variance (ANOVA). In the last step of the algorithm, the K-means segmentation method is used in the lesion block, to classify it as a tumor or not. A thorough evaluation of the proposed algorithm was performed, using: (1) a private MRI database provided by the Kouba imaging center—Algiers (KICA); (2) the multimodal brain tumor segmentation challenge (BraTS) 2015 database. Estimates of the selected fitness function were first compared to those based on the sum-of-absolute-differences (SAD) dissimilarity criterion, to demonstrate the efficiency and robustness of the ANOVA. The performance of the optimized brain tumor segmentation algorithm was then compared to the results of several state-of-the-art techniques. The results obtained, by using the Dice coefficient, Jaccard distance, correlation coefficient, and root mean square error (RMSE) measurements, demonstrated the superiority of the proposed optimized segmentation algorithm over equivalent techniques

International audience

International audience; Respiratory motion leads to blur and artifacts in single photon emission computed tomography (SPECT) images. This may be an issue in some treatments such as liver radioembolization. The purpose of this work was to reconstruct four-dimensional (4D) respiration-gated SPECT images using a data-driven approach. A respiratory signal was extracted from SPECT list-mode data using Laplacian Eigenmaps and used for retrospective phase gating. The method was evaluated on various SPECT datasets: Monte Carlo simulation of a patient acquisition, several acquisitions of a dynamic phantom and 20 acquisitions of patients treated with liver radioembolization. We observed an increase of the mean activity inside the tumor (A) and of the tumor-to-normal liver ratio (T/N) in respiration-gated SPECT image compared to the 3D image without motion correction. For the patient acquisitions, an average 3.3% (up to 11.9%) and 3.2% (up to 9.7%) increase was observed for A and T/N, respectively. For the simulation, the activity was within 2.6% of the reference without movement and around 3% for the three patterns of tumor motion simulated with the mechanical phantom. Overall, the proposed respiration-gated SPECT reconstruction improves the estimation of the activity in the tumor without the need for extra hardware.

The detection of new multiple sclerosis (MS) lesions is an important marker of the evolution of the disease. The applicability of learning-based methods could automate this task efficiently. However, the lack of annotated longitudinal data with new-appearing lesions is a limiting factor for the training of robust and generalizing models. In this work, we describe a deep-learning-based pipeline addressing the challenging task of detecting and segmenting new MS lesions. First, we propose to use transfer-learning from a model trained on a segmentation task using single time-points. Therefore, we exploit knowledge from an easier task and for which more annotated datasets are available. Second, we propose a data synthesis strategy to generate realistic longitudinal time-points with new lesions using single time-point scans. In this way, we pretrain our detection model on large synthetic annotated datasets. Finally, we use a data-augmentation technique designed to simulate data diversity in MRI. By doing that, we increase the size of the available small annotated longitudinal datasets. Our ablation study showed that each contribution lead to an enhancement of the segmentation accuracy. Using the proposed pipeline, we obtained the best score for the segmentation and the detection of new MS lesions in the MSSEG2 MICCAI challenge., preprint

International audience; In this study, we propose to use MR-Acoustic Radiation Force Imaging (MR-ARFI) as a transient elastography method to estimate the viscoelasticity and the anisotropy of an ultrasound absorbing medium.Methods: MR-ARFI1 uses the acoustic radiation force induced by a focused ultrasound beam to generate a remote micrometric displacement (push) at the ultrasonic focal region. This push results in the propagation of a transient shear wave of velocity cs that is encoded in the MRI phase signal by a motionsensitized sequence synchronized to the push. The resulting MR-ARFI phase profile depicts the ARFI spot, whose size increases with tissue stiffness. A previous quantitative MR-ARFI method2 has been proposed that allows estimating the local shear modulus μ of the medium around the focal spot, through the identification of the MR-ARFI profile. However, that work was limited to isotropic media, and the viscoelastic relaxation time τ was fixedarbitrarily. In this work, we propose to estimate τ following the approach published by Kaye et al.3 and Dadakova et al.4, which consists in fitting the phase signal at the focal spot for different temporal synchronization steps. Process is also extended to tissue anisotropy estimation, by performing individual fits on each radial profile extracted from the MR-ARFI phase image (Fig. A). The method is first validated on a homogeneous phantom calibrated for elastography (CIRS, USA; datasheet μref = 5.9 kPa ± 5%). Second, changes in biomechanical properties are estimated before and after a HIFU thermal ablation in ex vivo calf muscle tissue.Results: In the CIRS phantom, the value determined for τ is 3.3 ms at the focal spot. Using our MR-ARFI model, the polar representation of μ is isotropic, with an average value of 7.9 kPa ± 3% (Fig. B). In the calf tissue, τ decreases from 7.5 ms pre-HIFU to 4.5 ms post-HIFU, while μ globally decreases (Fig. C) suggesting that the meat softens as a result of heating. Figure C shows the shear modulus angular distribution both pre- and post-heating in calf tissues that appear to be highly anisotropic. The acquisition time for polar map of μ was 4 s, corresponding to the acquisitionof 2 MR images with opposite motion sensitivity polarities.Discussion: Quantitative MR-ARFI with viscoelastic model and anisotropic analysis provides shear modulus in the same range as μref in the CIRS phantom. After the HIFU experiment, cutting the calf tissue revealed a very irregular fiber orientation, consistent with the anisotropy observed with MR-ARFI. The maximum temperature rise reached in this experiment was 50 ºC, which results in meat tenderization, as already reported in the literature.Conclusion: The proposed method allows for local, quantitative, and direction-dependent viscoelasticity measurements around the focal spot. Its ability to provide polar maps of μ makes it a promising method to gain insights into biological tissue mechanical properties. The modification of τ found with HIFU heating calls for a further refined modeling of viscoelasticity to extend the method to the monitoring of HIFU ablations.

International audience; In order to build an intelligent dental care process that both facilitates the treatment and improves the diagnosis, an accurate tooth segmentation and recognition on panoramic X-ray images might prove helpful. Although many studies have been conducted on teeth segmentation, few methods allow to perform tooth recognition and numbering at the same time. The existing methods allowing both those processes rely on instance segmentation architectures. To fill some gaps in the area of dental image segmentation, we propose a novel approach of automatic joint teeth segmentation and numbering using the pioneer U-Net model. We are first to employ the conventional U-Net model and show its limitations to provide accurate segmentation, being affected by noisy pixels outside the teeth region and by missing teeth in the X-ray images. To overcome this problem and reduce the misclassifications, we use a bounding box prior at the level of the skip connections. Such an approach helps guiding the network to better locate the teeth, and hence improves the segmentation. To validate the effectiveness of the method, we have conducted two experiments on the DNS Panoramic Dataset: a first one using manual bounding boxes and another one relying on a preliminary step of object detection. The implemented networks were evaluated using the Dice coefficient index and our results showed that consideration of location information onto the skip connections improves the performances of the semantic segmentation by 5% to 10% in average Dice accuracy depending on the quality of the bounding box labels.

MRI has been extensively used to identify anatomical and functional differences in Autism Spectrum Disorder (ASD). Yet, many of these findings have proven difficult to replicate because studies rely on small cohorts and are built on many complex, undisclosed, analytic choices. We conducted an international challenge to predict ASD diagnosis from MRI data, where we provided preprocessed anatomical and functional MRI data from > 2,000 individuals. Evaluation of the predictions was rigorously blinded. 146 challengers submitted prediction algorithms, which were evaluated at the end of the challenge using unseen data and an additional acquisition site. On the best algorithms, we studied the importance of MRI modalities, brain regions, and sample size. We found evidence that MRI could predict ASD diagnosis: the 10 best algorithms reliably predicted diagnosis with AUC∼0.80 – far superior to what can be currently obtained using genotyping data in cohorts 20-times larger. We observed that functional MRI was more important for prediction than anatomical MRI, and that increasing sample size steadily increased prediction accuracy, providing an efficient strategy to improve biomarkers. We also observed that despite a strong incentive to generalise to unseen data, model development on a given dataset faces the risk of overfitting: performing well in cross-validation on the data at hand, but not generalising. Finally, we were able to predict ASD diagnosis on an external sample added after the end of the challenge (EU-AIMS), although with a lower prediction accuracy (AUC=0.72). This indicates that despite being based on a large multisite cohort, our challenge still produced biomarkers fragile in the face of dataset shifts.

International audience; Computational fluid dynamics (CFD) is a key tool for a wide range of research areas, beyond the computer science community. In particular, CFD is used in medicine to measure blood flow from patient specific models of arteries. In this field, the creation of accurate meshes remains the most challenging step, as it is based on the segmentation of medical images, a time-consuming task which often requires manual intervention by medical doctors. In this context, userfriendly, interactive softwares are valuable. They enable to spread the new advances in numerical treatment to the medical community and enrich them with the expert knowledge (e.g anatomical knowledge) of clinicians. In this work, we present a user interface dedicated to the meshing of vascular networks from centerlines. It allows for the 3D visualization and edition of input centerlines, which constitute a simplified, easy-tomanipulate representation of vascular networks. The surface of the artery can be reconstructed from the modified centerlines by an editable parametric model and then meshed with high quality hexahedral elements. At every step of the process, the network can be confronted with medical images with enhanced visualization. The software will be released publicly. Clinical relevance - This tool facilitates the manual extraction and editing of vascular networks by medical doctors. It opens the generation of hexahedral meshes for computational fluid dynamics studies to non-expert users.

Mapping the structural connectivity of the human brain is a major scientific challenge. Describing the trajectory and connections made by the hundred billion neurons that make up the brain is a titanic and multi-scale task.The major fiber bundles have been described by classical anatomical approaches since the 20th century. These studies also revealed the existence of shorter bundles, called superficial bundles, that ensure the connectivity between neighboring anatomical regions. The small size and complex shape of these bundles set a serious challenge to their visualization, so that their description remains under discussion to this day.The first research axis of this thesis aims at pushing the limits of diffusion MRI and proposing a new ex-vivo dataset of the whole human brain, called Chenonceau, dedicated to the characterization of the fine connectivity of the brain.The dataset consists of two T2-weighted anatomical acquisitions at 100 and 150 micron resolution, as well as 175 dMRI datasets at 200 micron resolution with diffusion weighting reaching 8000 s/mm2. More than 4500 hours of acquisition, distributed across two and a half years were necessary to acquire this data.Chenonceau takes advantage of the Bruker 11.7T preclinical MRI system, equipped with both a high magnetic field and a powerful gradient tunnel (780mT/m) allowing to reach the mesoscopic resolution and a very high diffusion weighting.To reconcile the large size of the human brain with the preclinical system, a new acquisition protocol is proposed. It is based on the separation of the brain into smaller samples, which are imaged individually, then reassembled in post-processing to reconstitute the full volume.The whole process is presented, including the protocol for the cutting and the storage of the anatomical samples, the details of the MRI sequences and the description of the image processing pipeline. Special attention is dedicated to the definition of the registration step which recomposes the whole volume from the individual acquisitions.The first inferences of anatomical connectivity from this new dataset are also presented. Tractography associated with clustering techniques allow the extraction of the long and superficial bundles of Chenonceau.The second part of the thesis focused on the development of a new method for fiber tracking, based on the use of the spin glass model.The latter expresses the tractography problem as a set of fiber fragments, called spins, distributed in the sample and whose position and orientation, as well as the connections they establish, are associated with an amount of energy. The construction of the tracts results from the displacement and connection of the spins, with the aim of reaching the global minimum of energy.This thesis proposes to replace the Metropolis-Hastings method used for optimization by an agent trained in a reinforcement learning framework.This new formulation aims at improving the choice of actions, which would no longer be randomly drawn, but dictated by a strategy learned by the agent, fruit of its past interactions with similar environments.The anticipation and projection capacities of such an agent appear particularly adequate to propose the most relevant trajectory in regions where the diffusion information is ambiguous. Moreover, the possibility for the algorithm to learn through interactions allows to circumvent the difficulty of establishing datasets of ground-truth bundles., La cartographie de la connectivité anatomique du cerveau humain est un défi scientifique majeur. Décrire la trajectoire et les connexions réalisées par les cent milliards de neurones qui composent le cerveau est une tâche titanesque et multi-échelle.Les grands faisceaux ont été décrits par des approches anatomiques classiques dès le 20ème siècle. Ces travaux ont également révélé l'existence de faisceaux plus courts, appelés superficiels, qui définissent la connectivité entre les régions anatomiques voisines. La taille réduite et la forme complexe de ces faisceaux posent un sérieux défi à leur visualisation, si bien que leur description demeure à ce jour débattue.Le premier axe de recherche de cette thèse vise à repousser les limites de l'IRM de diffusion et proposer un nouveau jeu de données ex-vivo du cerveau humain entier, intitulé Chenonceau, dédié à la caractérisation de la connectivité fine du cerveau.Le jeu de données est composé de deux acquisitions anatomiques pondérées en T2 à une résolution de 100 et 150 microns, ainsi que 175 jeux de données d'IRMd à une résolution de 200 microns et une pondération s'élevant jusqu'à 8000 s/mm2. Plus de 4500 heures d'acquisitions, réparties sur deux ans et demie ont été nécessaires pour acquérir ces données.Chenonceau met à profit la puissance de l'IRM pré-clinique Bruker 11.7T, doté à la fois d'un champ magnétique élevé et d'un tunnel de gradients puissants (780mT/m) permettant d'atteindre la résolution mésoscopique et une très forte pondération en diffusion.Pour concilier la taille imposante du cerveau humain avec l'imageur pré-clinique, un nouveau protocole d'acquisition est proposé. Celui-ci repose sur la séparation du cerveau en échantillons de taille réduite, qui sont sont imagés individuellement, puis réassemblés en post-traitement pour reconstituer le volume intégral.L'ensemble de la démarche est présenté, incluant le protocole de coupe et de préservation des pièces anatomiques, le détail des séquences IRM utilisées ainsi que la description du pipeline de traitement des images. Une attention particulière est portée à la définition de l'étape de recalage qui recompose le volume entier à partir des acquisitions individuelles.Les premières inférences de la connectivité anatomique issues de ce nouveau jeu de données sont également présentées. Les techniques de tractographie et de clustering permettent d'extraire non seulement les faisceaux longs de Chenonceau, mais également les faisceaux superficiels.La seconde partie de la thèse a porté sur le développement d'une nouvelle méthode de suivi de fibres, fondée sur l'utilisation d'un modèle de verres de spins.Ce dernier exprime le problème de tractographie sous la forme d'un ensemble de fragments de fibres, appelés spin, distribués dans l'échantillon et dont la position et l'orientation, ainsi que les connexions qu'ils établissent sont associés à une quantité d'énergie. La construction des tracts résulte du déplacement et de la connexion des spins, dans le but d'atteindre le minimum global d'énergie.Cette thèse propose de remplacer la méthode de Metropolis-Hastings utilisée pour l'optimisation par un agent entraîné dans un cadre d'apprentissage par renforcement.Cette nouvelle formulation vise à améliorer le choix des actions, qui ne seraient plus tirées aléatoirement, mais dictées par une stratégie apprise par l'agent, fruit de ses interactions passées avec des environnement semblables.Les capacités d'anticipation et de projection d'un tel agent apparaissent particulièrement adéquates pour proposer la trajectoire la plus pertinente dans des régions ou l'information de diffusion est ambiguë. De même, la possibilité pour l'algorithme d'apprendre au travers d'interactions permet de contourner la difficulté d'établir des ensembles de faisceaux considérées véritables.

In the human, the folding of the brain takes place mostly in utero, during the last trimester of pregnancy, resulting in a convoluted cortical surface at term, displaying furrows (sulci) and crests (gyri), whose patterns remain relatively stable throughout life. The same global folding organization is found between individuals, allowing to identify and compare matching sulci, while the more refined individual folding pattern is unique. The variability of sulcal folding pattern and its relations to functional specialization (both in health and pathology) are increasingly investigated in the mature brain, while studies in the developing brain are still scarce because of the many challenges of early-life neuroimaging. During my PhD, I have investigated the dynamics of folding pattern based on longitudinally acquired magnetic resonance imaging (MRI) images from a cohort of extremely preterm infants, imaged once around the equivalent of 7 months of pregnancy, and once at term-equivalent age (TEA). I have adapted a pipeline established for adult sulcal shape characterization to the context of developing folds and applied it to four sulcal objects: the central sulcus, the sylvian fissure, and the perisylvian region and the inferior frontal region (these last two were only investigated at TEA because of their late appearance in the brain).Throughout my works, I have described the shape variability observed in the central sulcus and the sylvian fissure at both ages. This led me to observe a consistent encoding of pattern between both ages, suggesting an early expression of sulcal shape features. I have incidentally tested if the central sulcus’s shape was able to predict later motor outcome: even though it did not achieve significant results, it still seemed to suggest some functional relevance in sulcal signatures. My observations on the central sulcus and sylvian fissure were concordant with the state of the art and detailed additional pattern specificities and the chronology of shape encoding. My focus in and around the sylvian fissure additionally led me to hypothesize a different dynamic in opercularization (the process leading to the burial of the insular cortex under the frontal, parietal and temporal lobes) from that in sulcation (the process leading to the formation of sulci): opercularization seemed to show an advance in the left hemisphere while sulcation has previously been reported to show an advance in the right hemisphere. Moreover, I have captured hemispheric asymmetries and detailed their shape characteristics in each fold-related object or region considered. These studies have therefore contributed to shedding more light on early-life asymmetries., which could be linked to the functional lateralization of the brain regarding handedness or language perception.In order to complement my contributions, additional investigations should be led in different directions, including the description of pattern variability and its links with brain function in later developmental stages in the child and the adult, the quantification of prematurity-induced sulcal alterations, and the further investigation of the relation between opercularization and sulcation processes, which might interact together during development., Chez l’homme, le cerveau se plisse principalement in utero, pendant le troisième trimestre de grossesse : la surface du cortex au moment de la naissance à terme présente alors des creux (sillons) et des bosses (gyri) dont la forme reste relativement stable au cours de la vie. Les motifs de plissement du cerveau, bien qu'uniques à chaque individu, présentent une organisation globalement cohérente au sein de l'espèce, permettant d'identifier et de comparer les sillons entre individus. La variabilité en forme des sillons et ses relations avec la spécialisation fonctionnelle (que le contexte soit pathologique ou non) sont de plus en plus étudiées dans le cerveau mature, mais les études dans le cerveau en développement restent rares à cause de nombreuses difficultés soulevées par la neuro-imagerie du tout-petit.Pendant mon doctorat, j’ai étudié les dynamiques de forme du plissement cortical sur la base de données d’imagerie par résonance magnétique (IRM) acquises longitudinalement dans une cohorte de nouveau-nés extrêmement prématurés, imagés vers l’âge équivalent à 7 mois de grossesse et à l’âge équivalent de la naissance à terme. J’ai adapté une méthode établie pour le cerveau adulte de caractérisation de la forme des sillons au contexte complexe du développement, et je l’ai appliquée à quatre régions corticales distinctes : le sillon central, la scissure sylvienne, et les régions périsylvienne et frontale inférieure, ces deux dernières n’ayant été explorées qu’à l’âge équivalent du terme du fait de leur apparition tardive dans le cerveau.À travers mes travaux, j’ai tout d’abord décrit la variabilité en forme du sillon central et de la scissure sylvienne. Ceci m’a amenée à découvrir que la forme de ces plis est cohérente entre les deux âges, ce qui suggère un encodage précoce. J’ai par ailleurs évalué si la forme du sillon central permettait de prédire le devenir moteur des sujets : faute d’être significative,elle a néanmoins suggéré la pertinence fonctionnelle de certaines signatures sulcales.Tout en étant concordantes avec l’état de l’art, mes observations sur le sillon central et la scissure sylvienne ont permis de détailler les motifs spécifiques et la chronologie d’encodage de leur forme précoce. En ciblant la scissure sylvienne et ses régions environnantes, j’ai émis l’hypothèse de dynamiques différentes entre l’opercularisation (processus menant au recouvrement du lobe insulaire par les lobes frontaux, pariétaux et temporaux) et la sulcation (processus menant à la formation des sillons) : mes observations ont alors suggéré une avance de l’hémisphère gauche dans le cadre de l’opercularisation, contre une avance de l’hémisphère droit pour la sulcation.Par ailleurs, j’ai capturé des asymétries conséquentes entre hémisphères cérébraux, en détaillant les caractéristiques de formes dans chacun des objets sulcaux considérés : ces travaux ont ainsi contribué à une meilleure compréhension des asymétries précoces qui pourraient être liées au développement de la latéralisation fonctionnelle du cerveau en liens avec la latéralité manuelle ou la perception du langage.De manière à compléter mes contributions, différentes directions devraient être explorées, dont la description de la variabilité des motifs des sillons et leurs relations avec le fonctionnement cérébral à des stades plus tardifs de développement de l’enfant et dans le cerveau de l’adulte, la quantification des altérations de plissement liées à la prématurité, et l’investigation plus approfondie des rapports entre opercularisation et sulcation, qui pourraient interagir ensemble au cours du développement.