Search

International audience; In this work, we address the problem of cone beam computed tomography (CBCT) image resolution enhancement by exploiting a newly introduced deep unrolled quantum denoiser, based on quantum interaction theory adapted to computational imaging. Following recent advances in image restoration using the Plug-and-Play (PnP) framework, we impose this external deep learning denoiser as a regularizer within the super-resolution (SR) problem. The quantum-based deep unrolled denoiser combined with a computationally efficient way to deal with the degradation operators, and the PnP scheme, result in an original way of approaching the image resolution enhancement problem. Experiments conducted on dental CBCT images are presented to illustrate the efficiency of the proposed model for image resolution enhancement tasks. The numerical results suggest that the proposed method allows better restoration performances compared to existing state-of-the-art approaches.

International audience; Two zinc finger peptides, namely ZFQD(Ln) and ZFQE(Ln) (Ln = Tb or Gd), with an appended Ln(3+) chelate and a phosphoserine able to coordinate the Ln(3+) ion are presented. The two peptides differ by the amino acid anchorage of the chelate, either aspartate (D) or glutamate (E). Both peptides are able to bind Zn2+ and adopt the beta beta alpha fold. Interestingly, ZFQE(Tb) shows a decrease in sensitized Tb3+ luminescence upon Zn2+ binding whereas ZFQD(Tb) does not. The luminescence change upon Zn2+ binding is attributed to a change in hydration number (q) of the Tb3+ ion due to the decoordination of the phosphoserine from the Ln(3+) ion upon Zn2+ binding and peptide folding. This process is highly sensitive to the length of the linker between the Ln chelate and the peptidic backbone. The magnetic properties of the gadolinium analogue ZFQE(Gd) were studied. An impressive relaxivity increase of 140% is observed at 60 MHz and 25 degrees C upon Zn2+ binding. These changes can be attributed to a combined increase effect of the hydration number of Gd3+ and of the rigidity of the system upon Zn2+ binding. Phantom MR images at 9.4 T show a clear signal enhancement in the presence of Zn2+. These zinc finger peptides offer a unique platform to design such Zn-responsive probes.

International audience

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; BACKGROUND: Medical diagnostic support systems are important tools in the field of radiology. However, the precision obtained, during the exploitation of high homogeneity image datasets, needs to be improved. OBJECTIVE: To develop a new learning system dedicated to public health practitioners. This study presents an upgraded version dedicated to radiology experts for better clinical decision-making when diagnosing and treating the patient (CAD approach). METHODS: Our system is a hybrid approach based on a matching of semantic and visual attributes of images. It is a combination of two complementary subsystems to form the intermodal system. The first one named α based on semantic attributes. Indexing and image retrieval based on specific key words. The second system named β based on low-level attributes. Vectors characterizing the digital content of the image (color, texture and shape) represent images. Our image database consists of 930 X-ray images including 320 mammograms acquired from the mini-MIAS database of mammograms and 610 X-rays acquired from the Public Hospital Establishment (EPH-Rouiba Algeria). The combination of two subsystems gives rise to the intermodal system: α-subsystem offers an overall result (based on visual descriptors), then β-subsystem (low level descriptors) refines the result and increases relevance. RESULTS: Our system can perform a specific image search (in a database of images with very high homogeneity) with an accuracy of around 90% for a recall of 25% . The average (overall) accuracy of the system exceeds 70% . CONCLUSION: The results obtained are very encouraging, and demonstrate efficiency of our approach, particularly for the intermodal system.

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; Purpose: The purpose of this study was to validate a new automated method to locate the fovea on normal and pathological fundus images. Compared to the normative anatomic measures (NAMs), our vessel-based fovea localization (VBFL) approach relies on the retina's vessel structure to make predictions. Methods: The spatial relationship between the fovea location and vessel characteristics is learnt from healthy fundus images and then used to predict fovea location in new images. We evaluate the VBFL method on three categories of fundus images: healthy images acquired with different head orientations and fixation locations, healthy images with simulated macular lesions, and pathological images from age-related macular degeneration (AMD). Results: For healthy images taken with the head tilted to the side, the NAM estimation error is significantly multiplied by 4, whereas VBFL yields no significant increase, representing a 73% reduction in prediction error. With simulated lesions, VBFL performance decreases significantly as lesion size increases and remains better than NAM until lesion size reaches 200 degrees 2. For pathological images, average prediction error was 2.8 degrees, with 64% of the images yielding an error of 2.5 degrees or less. VBFL was not robust for images showing darker regions and/or incomplete representation of the optic disk. Conclusions: The vascular structure provides enough information to precisely locate the fovea in fundus images in a way that is robust to head tilt, eccentric fixation location, missing vessels, and actual macular lesions. Translational Relevance: The VBFL method should allow researchers and clinicians to assess automatically the eccentricity of a newly developed area of fixation in fundus images with macular lesions.

Formalizing surgical activities as triplets of the used instruments, actions performed, and target anatomies is becoming a gold standard approach for surgical activity modeling. The benefit is that this formalization helps to obtain a more detailed understanding of tool-tissue interaction which can be used to develop better Artificial Intelligence assistance for image-guided surgery. Earlier efforts and the CholecTriplet challenge introduced in 2021 have put together techniques aimed at recognizing these triplets from surgical footage. Estimating also the spatial locations of the triplets would offer a more precise intraoperative context-aware decision support for computer-assisted intervention. This paper presents the CholecTriplet2022 challenge, which extends surgical action triplet modeling from recognition to detection. It includes weakly-supervised bounding box localization of every visible surgical instrument (or tool), as the key actors, and the modeling of each tool-activity in the form of triplet. The paper describes a baseline method and 10 new deep learning algorithms presented at the challenge to solve the task. It also provides thorough methodological comparisons of the methods, an in-depth analysis of the obtained results across multiple metrics, visual and procedural challenges; their significance, and useful insights for future research directions and applications in surgery., Comment: MICCAI EndoVis CholecTriplet2022 challenge report. Published at Elsevier journal of Medical Image Analysis. 25 pages, 15 figures, 8 tables

Accepted at 2023 IEEE 36th International Symposium on Computer-Based Medical Systems (CBMS).© 2023 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other work.; International audience; Multiple sclerosis (MS) patients often present with lesions in spinal cord magnetic resonance (MR) volumes. However, accurately detecting these lesions is challenging and prone to inter-and intra-rater variability. Deep learning-based methods have the potential to aid clinicians in detecting and segmenting MS lesions, but can also be affected by rater variability. This study assesses the inter-and intra-rater variability in manual segmentation of spinal cord lesions, and evaluates raters and a state-of-the-art nnU-Net model against a ground truth (GT) segmentation of a senior expert. Four experts segmented twelve spinal cord MR volumes from six patients twice, at a time distance of two weeks. Considerable inter-and intra-rater variability were observed, with the total number of detected lesions ranging from 28 to 60, depending on the rater. Moreover, the segmented volumes of individual lesions varied substantially between raters. All raters and the model achieved high precision when evaluated against the senior expert GT, but sensitivity was notably lower. These results motivate the need for more sensitive automated methods to aid clinicians in lesion detection, and suggest that consideration should be given to inter-rater variability when training and evaluating automated methods.

International audience; Cardiac Resynchronization Therapy (CRT) is a treatment aimed at restoring the electrical synchronization in patients with heart failure and intraventricular conduction delay. However, over 30% of patients do not respond to CRT. Septal Flash (SF), an abnormality characterized by a rapid inward-outward abnormal motion at early systole, has been linked to an improved response to CRT in patients with Left Bundle Branch Block (LBBB). In clinical practice, the detection of SF is usually performed manually through echocardiographic acquisitions, which is subjective and dependent on the operator's experience. To address this issue, a deep classification model for automatic SF detection from 2D anatomical M-mode images has been proposed. Additionally, this work focuses on SF localization from gradient-based attention maps, which provide a visual explanation of the output prediction of the model. Two models based on convolutional neural networks (CNNs) were trained with original and cropped M-modes from 143 patients, and achieved an accuracy of 0.83 and 1.0 respectively, on 29 testing patients. The attention map visualization showed that in SF cases, the models effectively identified the discriminant regions, while in non-SF cases, the maps appeared more dispersed. Further research is necessary to quantitatively evaluate the attention map results.

International audience; Measuring the thickness of a tissue can provide valuable clinical information; anatomical structures segmented on medical images can include sub-structures (inclusions) corresponding to a different biological tissue. This article presents a method, based on partial differential equations, to measure the thickness of one specific tissue in this particular configuration. After describing the mathematical formulation of our weighted thickness definition, we show on synthetic geometries in one, and two dimensions that it outputs the expected results. We then present three possible applications of our method on cardiac imaging data: measuring the muscular thickness of a ventricle with fat infiltration; measuring the thickness of an infarct scar; visualising the transmural extent of an infarct scar.

International audience; Echocardiography plays an important role in the diagnosis of cardiac dysfunction. In particular, motion estimation in echocardiography is a challenging task since ultrasound images suffer largely from low signal-to-noise ratio and out-of-view problems. Current deep learningbased models for cardiac motion estimation in the literature estimate the dense motion field with spatial regularization. However, the underlying spatial regularization can only cover a very small region in the neighborhood, which is not enough for a smooth and realistic motion field for the myocardium in echocardiography. In order to improve the performance and quality with deep learning networks, we propose applying polyaffine transformation for motion estimation, which intrinsically regularizes the myocardium motion to be polyaffine. Our thorough experiments demonstrate that the proposed method not only presents better evaluation metrics on the registration of cardiac structures but also shows great potential in abnormal wall motion detection.

The purpose of this thesis is to develop machine learning models that can leverage histology slides and clinical data to predict the outcome of immunotherapies against lung cancer. To this end, there are several challenges to overcome, such as the concurrent classification and localization of information within whole-slide images of large size, or the interpretability of the predictions made by the models. The thesis proposes several contributions to address such challenges.Chapter 2 introduces the concept of mixed supervision in histopathology. The purpose of this framework is to leverage several levels of supervision (i.e., global and local supervision) to make the model more efficient in both classification and localization tasks. Set on an attention-based deep learning model fit for global and local classification of tissue in whole-slide images, we show that it is possible to improve not only the model slide-level classification performance, but also and most importantly its ability to accurately locate regions of interest in the tissue, with only a few available local annotations.Chapter 3 extends the work presented in chapter 2, by consolidating the simultaneous classification and localization branches of the model with tailored loss functions that enforce attention distribution to follow the actual label distribution with respect to the available annotations. A slide sampling strategy is also proposed to strengthen the localization performance, and simplify the training procedure to have it fit in a single process.In Chapter 4, we present a multicentric lung cancer dataset dedicated to the prediction of immunotherapy response. We document the various steps followed to filter out low-quality samples, as well as undetermined outcomes, and we discuss the interpretation of what is a positive or a negative response with respect to the current gold standard clinical evaluations. Finally, we evaluate several models to directly predict the treatment response out of pathology slides, and discuss the caveats of the tested approaches.In Chapter 5, we switch from binary treatment response prediction to survival analysis, and use contrastive learning along with deep nonparametric clustering to generate a set of prognostic features in an unsupervised manner. We show that the obtained set of characteristics is a powerful indicator of survival, and that it maintains a good level of performance when picking one acquisition center as the test set. We also discuss the histological interpretation of the most prominent discovered clusters.To conclude, we address the remaining issues and challenges, and debate what future directions could be taken in order to tackle unanswered questions.; L’objectif de cette thèse est de développer des modèles d’apprentissage automatique capables d’exploiter des lames histologiques et des données cliniques pour prédire le résultat des traitements par immunothérapie contre le cancer du poumon. À cette fin, plusieurs défis doivent être relevés, tels que la classification et la localisation simultanées d’informations dans des images de lames entières de grande taille, ou l’interprétation des prédictions faites par les modèles. Dans ce qui suit, nous proposerons plusieurs contributions pour relever ces défis.Le chapitre 2 introduit le concept de supervision mélangée en histopathologie. L’objectif de cette méthode est de tirer parti de plusieurs niveaux de supervision (c’est-à-dire la supervision globale et locale) pour rendre le modèle plus efficace à la fois en classification et en localisation. Sur la base d’un modèle d’apprentissage profond basé sur l’attention et adapté à la classification globale et locale de tissu dans des coupes histologiques, nous montrons qu’il est possible d’améliorer non seulement les performances du modèle en matière de classification des lames, mais aussi et surtout sa capacité à localiser avec précision les régions d’intérêt dans le tissu disponible, lorsque seules quelques annotations disponibles.Le chapitre 3 étend le travail présenté dans le chapitre 2, en consolidant les branches de classification et de localisation simultanées du modèle avec des fonctions de coût adaptées qui contraignent la distribution de l’attention à suivre la distribution réelle des labels d’après les annotations disponibles. Une stratégie d’échantillonnage des images est également proposée pour renforcer les performances de localisation et simplifier la procédure d’apprentissage afin qu’elle s’inscrive dans un processus unique.Dans le chapitre 4, nous présentons un ensemble de données multicentriques sur le cancer du poumon dédié à la prédiction de la réponse à l’immunothérapie. Nous documentons les différentes étapes suivies pour éliminer les échantillons de faible qualité, ainsi que les cas indéterminés, et nous discutons de la définition de ce qu’est une réponse positive ou négative par rapport aux évaluations cliniques actuelles de référence. Enfin, nous évaluons plusieurs modèles permettant de prédire directement la réponse au traitement à partir de lames histologiques et discutons des écueils des approches envisagées. Dans le chapitre 5, nous passons de la prédiction binaire de la réponse au traitement à la prédiction de survie, et nous utilisons l’apprentissage par contraste ainsi que le regroupement non paramétrique profond pour générer un ensemble de caractéristiques pronostiques de manière non supervisée. Nous montrons que l’ensemble des caractéristiques obtenu est un puissant indicateur de survie et qu’il conserve un bon niveau de performance lorsque l’on choisit un seul centre comme ensemble de test. Nous discutons également de l’interprétation histologique faite du résultat de l’algorithme de regroupement, en particulier pour les groupes les plus corrélés à la survie.Pour conclure, nous abordons les questions et les défis en suspens, et nous discutons des orientations futures qui pourraient être prises afin de répondre aux questions restées sans réponse.

Surgical video analysis facilitates education and research. However, video recordings of endoscopic surgeries can contain privacy-sensitive information, especially if the endoscopic camera is moved out of the body of patients and out-of-body scenes are recorded. Therefore, identification of out-of-body scenes in endoscopic videos is of major importance to preserve the privacy of patients and operating room staff. This study developed and validated a deep learning model for the identification of out-of-body images in endoscopic videos. The model was trained and evaluated on an internal dataset of 12 different types of laparoscopic and robotic surgeries and was externally validated on two independent multicentric test datasets of laparoscopic gastric bypass and cholecystectomy surgeries. Model performance was evaluated compared to human ground truth annotations measuring the receiver operating characteristic area under the curve (ROC AUC). The internal dataset consisting of 356,267 images from 48 videos and the two multicentric test datasets consisting of 54,385 and 58,349 images from 10 and 20 videos, respectively, were annotated. The model identified out-of-body images with 99.97% ROC AUC on the internal test dataset. Mean ± standard deviation ROC AUC on the multicentric gastric bypass dataset was 99.94 ± 0.07% and 99.71 ± 0.40% on the multicentric cholecystectomy dataset, respectively. The model can reliably identify out-of-body images in endoscopic videos and is publicly shared. This facilitates privacy preservation in surgical video analysis.

International audience; Reproducibility is a cornerstone of science, as the replication of findings is the process through which they become knowledge. It is widely considered that many fields of science are undergoing a reproducibility crisis. This has led to the publications of various guidelines in order to improve research reproducibility. This didactic chapter intends at being an introduction to reproducibility for researchers in the field of machine learning for medical imaging. We first distinguish between different types of reproducibility. For each of them, we aim at defining it, at describing the requirements to achieve it and at discussing its utility. The chapter ends with a discussion on the benefits of reproducibility and with a plea for a non-dogmatic approach to this concept and its implementation in research practice.

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.

International audience; We present a new method to automatically identify the different arteries present in an abdominal aortic segmentation. In this approach, the arterial system is first represented by a vascular tree, extracted from the segmentation and containing the topologic and geometric features (branch position, branch direction, branch length, branch diameter) of the arterial system. Then, the branches of the vascular tree are matched with the main arteries origi- nating from the aorta: celiac artery, superior mesenteric artery, renal arteries and common iliac arteries. This match is determined by maximizing a similarity measure between the dif- ferent branches and corresponding arteries. We evaluate this method on 239 segmentations obtained from 102 different patients. The results demonstrate the accuracy of the proposed method, capable of delivering an error of less than 2.5% for the identification of the celiac and superior mesenteric arteries, 8.4% for the renal arteries, and 2.1% for the common iliac arteries.

Magnetic Resonance Imaging (MRI) pulse sequence programming relies on development kits provided by manufacturers and allows for the design of new acquisition methods. However, due to intellectual property issues, the sharing of these sequences for research purposes is restrained, thus hindering their development. To overcome these difficulties, researchers proposed vendor-independent solutions for the prototyping of MRI pulse sequences. However, as these environments rely on high-level interfaces, their use for accessing and operating the user interface optimally or for performing real-time protocols is not straightforward. Allowing for the open-source development of MRI pulse sequences capable of using all the complex features proposed by the manufacturer would thus support the development of new pulse sequences for research purposes. Proposing a solution to this issue, this thesis presents the development of GinkgoSequence, a novel modular and open-source MRI pulse sequence programming framework for Siemens systems. The first part of this work details the principles of nuclear magnetic resonance and MRI, specifying the notions of spatial encoding and pulse sequence. It presents several methods for rapid acquisition, including echo-planar imaging and accelerated methods relying on partial data acquisition; before introducing tools for MRI pulse sequence development and image reconstruction, with a description of Siemens’s IDEA development kit as well as a panel of existing open-source solutions for pulse sequence programming. It also introduces the applications of MRI for human brain imaging, starting with a brief anatomical description before presenting the anatomic, functional, and diffusion-weighted MRI modalities. The second part of this thesis focuses on the development of the GinkgoSequence toolbox, defining its challenges and technical requirements and proving its robustness and efficiency by providing example acquisitions on phantoms and human volunteers. First, it proposes a detailed description of its object-oriented source code based on sequence modules standing for fundamental parts of any MRI pulse sequence that can be ordered using an- chor times. These modules are embedded into a factory and can be aggregated into composite modules before being organized in time to form a pulse sequences source code that can be compiled using the IDEA environment. The toolbox includes examples of sequences compatible with the open- source Gadgetron framework for reconstruction. Second, it details the development of standard sequences, showing and analyzing images acquired on a phantom and comparing them with reference sequences. These trials prove the robustness of the basic modules provided in the GinkgoSequence toolbox and their ability to obtain results conforming with the images acquired using reference sequences. Third, it describes the further development of the toolbox to implement a series of diffusion- weighted MRI pulse sequences and evaluates their results on a few healthy volunteers for whom a reconstruction of the structural connectivity and quantitative microstructure was successfully achieved. Last, it introduces the implementation of an echo-volumar sequence, allowing an ultra-fast acquisition that can be exploited both for diffusion and functional MRI, and its use to scan an ex-vivo human hemisphere, assessing the quality of the re- constructed data. In summary, after presenting its scientific context and challenges, this thesis details the development of a toolbox allowing for the open-source and modular development of MRI pulse sequences. It demonstrates the robustness of its basis by providing images acquired with standard sequences and its capacity to efficiently produce pulses sequences that can be used in a research environment with the given examples of diffusion MRI and echo-volumar imaging.; La programmation de séquences d’Imagerie Par Résonance Magnétique (IRM) repose sur l’utilisation d’environnements de développements fournis par les constructeurs. Ces derniers permettent l’élaboration de nouvelles méthodes d’acquisition mais sont limités par des questions de propriété intellectuelle. Des solutions ont été proposées par la communauté scientifique permettant le prototypage de séquences indépendamment du constructeur. Cependant, ils utilisent une interface haut-niveau ne facilitant pas l’accès à toute les fonctionnalités des machines, ce qui restreint le développement de nouvelles séquences en ne permettant pas l’utilisation optimale de l’interface utilisateur, ni le déploiement de séquences optimisées pour le temps réel. L’ouverture de ces possibilités présente un potentiel considérable pour la recherche en IRM et n’est pour le moment possible que dans les conditions imposées par le constructeur. Afin de proposer une solution à ce problème, cette thèse présente le développement de GinkgoSequence, un nouvel environnement libre de droits pour la programmation modulaire de séquences dans l’environnement Siemens. La première partie présente les principes de la résonance magnétique nucléaire et de l’IRM, détaillant les notions d’encodage spatial et de séquence. Elle décrit les méthodes d’acquisition rapides, incluant l’imagerie écho-planaire et les techniques d’accélération reposant sur l’acquisition partielle des données, ainsi que les outils pour la programmation de séquences et la reconstruction des images, avec une description du kit IDEA de Siemens ainsi que des solutions libres de droits existantes. Elle introduit ensuite l’application de l’IRM à l’imagerie du cerveau humain, présentant l’anatomie de ce dernier avant de détailler les méthodes d’IRM anatomique, fonctionnelle, et pondérée en diffusion. La deuxième partie se focalise sur le développement de GinkgoSequence, ses enjeux et modalités, avant de prouver sa robustesse et son efficacité à l’aide d’acquisitions sur fantôme et sur volontaire sain. Dans un premier temps, elle décrit son contenu orienté-objet basé sur la notion de modules représentant les parties élémentaires d’une séquence IRM et pouvant être agencés à l’aide d’ancres temporelles. Ces modules sont générés en utilisant une factory et peuvent être agrégés dans des modules composites avant d’être organisés dans le temps pour former une séquence compilable avec IDEA. La toolbox inclut des exemples de séquences dont les images peuvent être reconstruites à l’aide de l’environnement open-source Gadgetron. Dans un second temps, elle présente des séquences de routine, montrant et analysant des images acquises sur un fantôme avant de les comparer avec l’existant. Ces essais prouvent la robustesse des modules de base et leur capacité à obtenir des résultats équivalents à ceux de séquences de référence. Dans un troisième temps, elle introduit l’utilisation de la toolbox pour l’implémentation de séquences IRM pondérés en diffusion et leur utilisation sur quelques volontaires sains pour évaluer les résultats des cartes de connectivité structurelle et de description quantitative de la microstructure obtenues avec succès. Enfin, elle aborde le développement d’une séquence écho-volumaire qui pourra être exploitée ans le cadre de séquences IRM fonctionnelles ou de diffusion. Elle montre l’utilisation de cette séquence pour scanner un hémisphère de cerveau humain ex vivo et évaluer la qualité des images reconstruites. En conclusion, après avoir présenté son contexte scientifique et ses enjeux, cette thèse présente GinkgoSequence, un outil pour le développement libre de droits et modulaire de séquences IRM. Elle démontre sa robustesse en présentant des images acquises avec des séquences des routine clinique et sa capacité à produire de nouvelles séquences avec les exemples donnés de l’IRM de diffusion et des séquences écho-volumaires.

Advancements in 7T MRI have allowed for higher spatial image resolution and better signal-to-noise ratio for the visualization of small vessels in the brain. Given the current availability of data, most segmentation techniques in the field have focused on large blood vessel segmentation as opposed to small blood vessel extraction. Thus, there is a need to train effective small vessel segmentation tools with minimal data. In this paper, we propose a scrolling 2D U-Net that integrates information regarding the 3D volume while performing 2D segmentation. This approach was compared to the nnUnet – the state-of-the-art in segmentation models.

International audience

Magnetic resonance Imaging (MRI) is a popular non-invasive medical imaging technique to understand structures of tissues in the human body. However, its reach and use is strongly limited by long acquisition times, as the data acquisition is segmented in time and performed in the Fourier domain of the image or k-space. Compressed sensing theories enabled shorter scan times by sensing less compressible information, particularly by adopting a variable density sampling (VDS) in k-space: i.e. the lower frequencies (center of k-space) have to be sampled more densely than the higher at its periphery. Efficient ways to achieve such VDS is through the use of non-Cartesian sampling, where k-space is sampled along curves rather than conventional gridded lines. Essentially, this involves the optimization of 1) the sampling pattern in k-space under MR hardware constraints and 2) image reconstruction from undersampled k-space data.In this thesis, we address the above issues in the context of accelerated 3D MRI. Firstly, we extend the earlier developed SPARKLING algorithm to 3D, which results in hardware compliant non-Cartesian trajectories that meet a user specified target sampling density (TSD) in k-space. Secondly, we tackle the issue of off-resonance artifacts in these SPARKLING trajectories through a modification of the underlying algorithm, to allow for signal recovery at regions where the off-resonance is high. Thirdly, we shift to data-driven learning based approach to learn the TSD for SPARKLING trajectories. Last but not least, we propose a novel deep learning based approach to learn hardware compliant k-space sampling trajectories jointly with a corresponding image reconstruction network. Overall, our efforts in this thesis allowed for an acceleration of up to 20x in 3D MRI scans with a significant reduction in off-resonance artifacts, allowing to image with minimum degradation in reconstructed image quality.; L'imagerie par résonance magnétique (IRM) est une technique d'imagerie médicale non invasive très répandue qui permet de comprendre les structures des tissus du corps humain. Cependant, sa portée et son utilisation sont fortement limitées par les longs temps d'acquisition, car l'acquisition des données est segmentée en temps et réalisée dans le domaine de Fourier de l'image ou espace k. Les théories d'échantillonnage compressif ont permis d'accéder à des temps d'examen plus courts en collectant moins d'information compressible, en particulier en adoptant d’un échantillonnage à densité variable (EDV) dans l'espace k : les basses fréquences (au centre de l'espace k) doivent être échantillonnées plus densément que les hautes fréquences situées à la périphérie de celui-ci. Un moyen efficace d'obtenir un tel EDV est d'utiliser un échantillonnage non cartésien, où l'espace k est échantillonneur le long de courbes plutôt que de lignes alignées sur une grille cartésienne. Essentiellement, cela implique l'optimisation 1) du modèle d'échantillonnage dans l'espace k sous les contraintes matérielles de l'IRM et 2) la reconstruction de l'image à partir des données sous-échantillonnées de l'espace k. Dans cette thèse, nous abordons les questions ci-dessus dans le contexte de l'IRM 3D accélérée. Tout d'abord, nous étendons l'algorithme SPARKLING développé précédemment à l'imagerie 3D, ce qui permet d'obtenir des trajectoires non cartésiennes conformes aux contraintes matérielles et fidèles à une densité d'échantillonnage cible (DEC) spécifiée par l'utilisateur dans l'espace k. Ensuite, nous abordons le problème des artefacts hors résonance présents au sein des trajectoires SPARKLING en modifiant l'algorithme sous-jacent, afin de permettre la récupération du signal dans les régions où la hors résonance est élevée. Dans la dernière partie de la thèse, nous passons à une approche basée sur l'apprentissage piloté par les données pour apprendre la DEC pour les trajectoires SPARKLING. Enfin, nous proposons une nouvelle approche basée sur l'apprentissage profond pour apprendre des courbes d'échantillonnage de l'espace k conformes aux contraintes matérielles conjointement avec un réseau de reconstruction d'images adapté. Dans l'ensemble, nos efforts dans cette thèse ont permis d'accélérer jusqu'à 20 fois les scans IRM 3D avec une réduction significative des artefacts hors résonance, permettant ainsi une dégradation minimale de la qualité des images reconstruites.

International audience; Age, sex and APOE-ε4 genotype have been identified as the strongest predictors of the risk of developing Alzheimer's Disease (AD). This work models the pathological progression of regional brain hypometabolism, using mixed-effect models with latent time variable and longitudinal FDG-PET data. Statistical comparisons then disentangle the effects of sex and APOE-ε4 genotype on the onset age and pace of progression of hympometabolism in each brain region, while correcting for education level. They provide a brain map of the regions with earlier and/or faster alterations of the metabolism. We show that females are associated with faster hypometabolism in the caudate nuclei, the thalamus and right temporal and medial-occipital lobes, while APOE-ε4 is associated with earlier hypometabolism in the limbic system (hippocampus, parahippocampus and amygdala) and temporal lobe.

International audience; An important issue in medical image processing is to be able to estimate not only the performances of algorithms but also the precision of the estimation of these performances. Reporting precision typically amounts to reporting standard-error of the mean (SEM) or equivalently confidence intervals. However, this is rarely done in medical image segmentation studies. In this paper, we aim to estimate what is the typical confidence that can be expected in such studies. To that end, we first perform experiments for Dice metric estimation using a standard deep learning model (U-net) and a classical task from the Medical Segmentation Decathlon. We extensively study precision estimation using both Gaussian assumption and bootstrapping (which does not require any assumption on the distribution). We then perform simulations for other test set sizes and performance spreads. Overall, our work shows that small test sets lead to wide confidence intervals (e.g. ∼8 points of Dice for 20 samples with σ ≃ 10).

International audience

International audience; Neural network-based anomaly detection remains challenging in clinical applications with little or no supervised information and subtle anomalies such as hardly visible brain lesions. Among unsupervised methods, patch-based auto-encoders with their efficient representation power provided by their latent space, have shown good results for visible lesion detection. However, the commonly used reconstruction error criterion may limit their performance when facing less obvious lesions. In this work, we design two alternative detection criteria. They are derived from multivariate analysis and can more directly capture information from latent space representations. Their performance compares favorably with two additional supervised learning methods, on a difficult de novo Parkinson Disease (PD) classification task.

International audience; Functional Magnetic Resonance Imaging (fMRI) data provides deep insight on brain activity, but high-resolution (e.g. 1 mm isotropic) fMRI suffers from low signal-to-noise ratio (SNR). Recently several denoising methods have been proposed to boost the SNR of high-resolution fMRI images. In this work we carry out a prospective benchmark of local low-rank denoising methods by quantifying their performances on a retinotopy experimental paradigm that was submitted to six healthy volunteers during a 7 Tesla, 1 mm isotropic fMRI acquisition protocol. Out of the five tested approaches (NORDIC, MP-PCA, Hybrid-PCA, Optimal Threshold (OT) and Hybrid-OT), the OT approach outperforms its competitors in terms of statistical sensitivity and specificity both at the subject-and group-level. Additionally, by testing these denoising methods in different configurations of preprocessing pipelines, we demonstrate first that on average it is beneficial to denoise fMRI images prior to performing realignment (i.e. motion correction), and second that the OT approach performs better when applied to complex-valued fMRI images instead of magnitude-only ones. We also provide an open source implementation to promote a broader use of denoising methods in fMRI and enable reliable statistical data analysis at high spatial resolution.

International audience; As a long-term consequence of neonatal arterial ischaemic stroke (NAIS), the presence of cerebral palsy (CP) depends on the structural integrity of brain areas, especially of basal ganglia. Yet, it remains challenging to establish an early diagnosis of CP from a conventional structural MRI. In this study, we introduce a graph neural network-based classification for the recognition of NAIS children and mainly for the detection of children with CP among the NAIS ones. From the structural MRI of 68 children aged 7 years old and their corresponding segmentation of basal ganglia, we construct graphs where nodes represent structures, carrying on node and edge attributes structural information (volumes, distances). The classification accuracy achieved by the proposed method is of 86% for the detection of NAIS and of 89% for the detection of CP among neonatal stroke children.

Dynamic contrast-enhanced (DCE) perfusion imaging has shown great potential to non-invasively assess cancer development and its treatment by their characteristic tissue signatures. Different tracer kinetics models are being applied to estimate tissue and tumor perfusion parameters from DCE perfusion imaging. The goal of this work is to provide an in silico model-based pipeline to evaluate how these DCE imaging parameters may relate to the true tissue parameters. As histology data provides detailed microstructural but not functional parameters, this work can also help to better interpret such data. To this aim in silico vasculatures are constructed and the spread of contrast agent in the tissue is simulated. As a proof of principle we show the evaluation procedure of two tracer kinetic models from in silico contrast-agent perfusion data after a bolus injection. Representative microvascular arterial and venous trees are constructed in silico. Blood flow is computed in the different vessels. Contrast-agent input in the feeding artery, intra-vascular transport, intra-extravascular exchange and diffusion within the interstitial space are modeled. From this spatiotemporal model, intensity maps are computed leading to in silico dynamic perfusion images. Various tumor vascularizations (architecture and function) are studied and show spatiotemporal contrast imaging dynamics characteristic of in vivo tumor morphotypes. The Brix II also called 2CXM, and extended Tofts tracer-kinetics models common in DCE imaging are then applied to recover perfusion parameters that are compared with the ground truth parameters of the in silico spatiotemporal models. The results show that tumor features can be well identified for a certain permeability range. The simulation results in this work indicate that taking into account space explicitly to estimate perfusion parameters may lead to significant improvements in the perfusion interpretation of the current tracer-kinetics models.

International audience; Narrow-band imaging (NBI) bronchoscopy offers enhanced visualization of microvascular structures in the lung’s epithelium (airway walls). Recent studies suggest that such vessels are helpful in predicting the invasiveness of bronchial lesions. In particular, Shibuya characterized pathological features of lesions and studied their relationship with specific histological stages of lung cancer. We propose a method for identifying these vascular patterns using a small expert-labeled dataset. Our approach is based on a few-shot learning method using a Siamese network to learn and distinguish pathological features of the bronchial vasculature. We achieved better intra-class clustering and inter-class separation in the embedding space compared to a baseline CNN classifier. Further, a 25% increase in the overall accuracy was obtained during testing.

One of the visual characteristics of the brain is its folded appearance. It is composed of convolutions called gyri, which are delimited by furrows called sulci. Like fingerprints, the patterns of cortical folding, i.e. the arrangement, characteristics and shape of the sulci, are unique to each individual, although they do fit into a general organization in each species.Some patterns have been correlated with cognitive functions such as inhibitory control in the cingulate region. In addition, rare patterns have been linked with neurodevelopmental disorders such as the "Power Button Sign" which is associated with a certain type of epilepsy. Advances in deep and especially unsupervised learning provide a real opportunity to analyze cortical folding patterns, addressing the challenge of dealing with large databases and high inter-individual variability.The aim of this thesis is to develop a method to identify rare or abnormal patterns. From brain MRI, skeletons and distance maps corresponding to a negative cast of the brain are generated. They allow the deep network to focus on the folding morphology. Inter-individual variability is modeled by a beta-VAE trained only on control subjects from the HCP cohort. From this normal representation, typical patterns of the healthy population can be identified. This first study is done in the cingulate region. More specifically, clustering is applied to the latent space. Representative patterns of each of the clusters are generated and allow us to characterize the different typical patterns. Patterns described in the literature were found, which suggests that the approach is relevant. Once characteristic patterns have been identified, rare patterns can be analyzed. In this case, the network is applied in the central sulcus area to cohorts with healthy subjects and subjects with rare patterns or to synthetic data with anomalies. The latter allows for a better qualification of the identified deviations. The identification of rare patterns is done in the latent space and on the basis of the reconstruction error. To evaluate the generalization, this approach is applied to two other regions with abnormal patterns. The results showed that both the latent space and the reconstruction errors provided complementary information and were more suitable for certain types of deviations. The generative power of the beta-VAE allows a better understanding of the properties encoded in the latent space. The visualization of reconstruction errors helps to localize the identified atypical features. Finally, the application of the method to the two other cohorts indicates a good generalization, despite a slight site effect. These promising results need to be replicated with more control subjects and subjects with rare patterns in order to be confirmed. The approach we propose could thus constitute a first step towards the identification of biomarkers and towards a systematic model at the brain scale. Finally, the proposed approach would benefit from the integration of other modalities such as cytoarchitecture or connectivity.; Une des caractéristiques visuelles du cerveau est son aspect plissé. Il est en effet constitué de circonvolutions appelées gyri, qui sont délimitées par des sillons appelés sulci. À l'instar des empreintes digitales, les motifs du plissement cortical, c'est-à-dire l'arrangement, les caractéristiques et la forme des sillons sont propres à chaque individu, même s'ils s'inscrivent dans un schéma général propre à chaque espèce. Certains motifs ont été corrélés à des fonctions cognitives comme le contrôle inhibiteur dans la région cingulaire. De plus, des motifs rares ont été corrélés à des troubles neuro-développementaux comme le « Power Button Sign » qui est associé à un certain type d'épilepsie. Les avancées de l'apprentissage profond et en particulier non supervisé constituent une réelle opportunité pour analyser les motifs du plissement cortical, répondant au défi des grandes bases de données et de la variabilité inter-individuelle. Ce travail de thèse a pour but de développer une méthode permettant d'identifier des motifs rares ou anormaux. À partir d'IRM cérébrales, des squelettes et des cartes de distances correspondant à un moule en négatif du cerveau sont générés. Ils permettent de se concentrer sur la morphologie du plissement. La variabilité inter-individuelle est modélisée par un beta-VAE entraîné uniquement sur des sujets témoins de la cohorte HCP. À partir de cette représentation, des motifs caractéristiques de la population saine peuvent être identifiés. La première étude est faite dans la région cingulaire. Plus spécifiquement un clustering est appliqué sur l'espace latent. Des motifs représentatifs de chacun des clusters sont générés et permettent de caractériser les différents motifs caractéristiques. Des motifs décrits dans la littérature ont été retrouvés ce qui suggère que l'approche est pertinente. Une fois des motifs caractéristiques identifiés, les motifs plus rares peuvent être analysés. Dans ce cas, le réseau est appliqué dans la région du sillon central à des cohortes comportant des sujets sains et des sujets avec des motifs rares ou à des données synthétiques présentant des anomalies. Ces dernières permettent de mieux qualifier les déviations identifiées. L'identification de motifs rares se fait dans l'espace latent ou sur la base de l'erreur de reconstruction. Pour évaluer la généralisation, cette approche est appliquée à deux autres régions comportant des motifs anormaux. Les résultats ont montré que l'espace latent et les erreurs de reconstruction apportent des informations complémentaires, adaptées à divers types de déviations. Le pouvoir génératif du beta-VAE permet de mieux comprendre les propriétés encodées dans l'espace latent. La visualisation des erreurs de reconstructions aide à localiser les caractéristiques atypiques identifiées. Enfin, l'application de la méthode aux deux autres cohortes indique une bonne généralisation, malgré un léger effet site. Ces résultats prometteurs demandent à être répliqués avec davantage de sujets contrôles et de sujets présentant des motifs rares afin d'être confirmés. L'approche que nous proposons pourrait ainsi constituer une première étape vers l'identification de biomarqueurs et vers un modèle systématique à l'échelle du cerveau. Enfin, l'intégration d'autres modalités comme la cytoarchitectonie ou la connectivité enrichirait l'approche proposée.

International audience

International audience; Accurate lung nodule localization during video-assisted thoracic surgery (VATS) for the treatment of early stage lung cancer is a surgical challenge. Recently, a new minimally invasive approach for nodule localization during VATS has been proposed, which consists in compensating by a biomechanical model the very large lung deformations occurring before and during surgery. This estimation of the deformations allows to transfer the position of the nodule visible on the preoperative CT to an acquisition of the lung performed during the operation using a Cone-Beam CT scanner (CBCT 2). But, in this approach, an additional CBCT acquisition (CBCT 1) must also be acquired just after the patient is placed in the operative position in order to estimate the deformations due to the change of the patient's position, from supine during the CT acquisition to lateral decubitus in the operating room. Our goal is to simplify this procedure and thus reduce the radiation dose to the patient. To this end, we propose to improve this solution by replacing the CBCT 1 acquisition by a model allowing to predict these deformations. This model is defined using the lung state information from CBCT 2 and a general statistical motion model built from the position change deformations already observed on other patients. We have data from 17 patients. The method is evaluated with a leave-one-out cross-validation on its ability to reproduce the observed deformations. The method reduces the average prediction error from 12.14 mm without prediction to 8.12 mm for an average prediction, and finally to 6.33 mm for a prediction with our model fitted to CBCT 2 only.

International audience; As deep learning has been widely used for computer aided-diagnosis, we wished to know whether attribution maps obtained using gradient back-propagation could correctly highlight the patterns of disease subtypes discovered by a deep learning classifier. As the correctness of attribution maps is difficult to evaluate directly on medical images, we used synthetic data mimicking the difference between brain MRI of controls and demented patients to design more reliable evaluation criteria of attribution maps. We demonstrated that attribution maps may mix the regions associated with different subtypes for small data sets while they could accurately characterize both subtypes using a large data set. We then proposed simple data augmentation techniques and showed that they could improve the coherence of the explanations for a small data set.

International audience

International audience; INTRODUCTIONThe Centiloid scale aims to harmonize amyloid beta (Aβ) positron emission tomography (PET) measures across different analysis methods. As Centiloids were created using PET/computerized tomography (CT) data and are influenced by scanner differences, we investigated the Centiloid transformation with data from Insight 46 acquired with PET/magnetic resonanceimaging (MRI).METHODSWe transformed standardized uptake value ratios (SUVRs) from 432 florbetapir PET/MRI scans processed using whole cerebellum (WC) and white matter (WM) references, with and without partial volume correction. Gaussian-mixture-modelling–derived cutpoints for Aβ PET positivity were converted.RESULTSThe Centiloid cutpoint was 14.2 for WC SUVRs. The relationship between WM and WC uptake differed between the calibration and testing datasets, producing implausibly low WM-based Centiloids. Linear adjustment produced a WM-based cutpoint of 18.1.DISCUSSIONTransformation of PET/MRI florbetapir data to Centiloids is valid. However, further understanding of the effects of acquisition or biological factors on the transformation using a WM reference is needed.HIGHLIGHTS Centiloid conversion of amyloid beta positron emission tomography (PET) data aims to standardize results. Centiloid values can be influenced by differences in acquisition. We converted florbetapir PET/magnetic resonance imaging data from a large birth cohort. Whole cerebellum referenced values could be reliably transformed to Centiloids. White matter referenced values may be less generalizable between datasets.

International audience

International audience; Evaluer la réponse thérapeutique tumorale : intérêt d’une approche bimodale en IRM CEST et imagerie nucléaire. Roxane Autissier1,2,3, Guilhem Pagés2,3, Erwan Boutault1, Jean-Marie Bonny2,3, Sébastien Schmitt1, Philippe Auzeloux1, Emmanuel Moreau1, Aurélie Dutour4, Elisabeth Miot-Noirault1, Leslie Mazuel1,2,3.1 Université Clermont Auvergne, Inserm, IMoST, F-63000 Clermont-Ferrand, France INRAe, 2 UR QuaPA, F-63122 Saint-Genès-Champanelle, France3AgroResonance, INRAe, 2018. Nuclear Magnetic Resonance Facility for agronomy, food and health, https://doi.org/10.15454/1.5572398324758228E123,44Team Cell Death and Pediatric Cancer, Cancer Initiation and Tumor Cell Identity Department, INSERM1052, CNRS5286, Cancer Research Center of Lyon, F-69008 Lyon, FranceObjectifs : Le chondrosarcome (CHS), tumeur maligne du cartilage, dispose d’un microenvironnement caractérisé par une matrice extracellulaire (MEC) riche en glycosaminoglycanes (GAGs) et par un cœur hypoxique. Ces deux caractéristiques sont identifiables par une stratégie d’acquisitions bimodales par IRM CEST et imagerie nucléaire [1]. Dans ce contexte, nous souhaitons évaluer la pertinence de cette stratégie pour déceler les modifications ayant lieu au sein de la tumeur en réponse à une thérapie ciblant la MEC (acide zolédronique) ou l’hypoxie (TH-302). Matériels et Méthodes : Les rats SD OFA implantés en paratibial avec un fragment de CHS (SWARM) sont traités ensuite par 2 injections/semaine d’acide zolédronique (0,1mg/kg, sc, J5-J36) ou 5 injections/semaine de TH-302 (Evofosphamide) (50mg/kg, ip, J11-J22) ou de sérum physiologique. Les animaux sont imagés hebdomadairement jusqu’à J40 par IRM à 11,7T. Un repérage anatomique puis une séquence RARE comprenant un module de préparation CEST (B1=1,5µT, tsat=4 s, Δω=50Hz sur  2500 Hz) est réalisée sur une coupe axiale passant par le centre de la tumeur. Une carte B0 est enregistrée à l’aide d’une WASSR (B1=0,1µT, tsat=1s, Δω=20Hz sur 1000Hz). Des imageries TEMP (99mTc-NTP15-5, 30MBq/animal, 30min pi) et TEP (18F-FMISO, 20MBq/animal, 4h pi) sont réalisées sur les mêmes animaux pour caractériser les GAGs et l’hypoxie, respectivement. Les tumeurs prélevées sont étudiées par histologie, immunofluorescence et dosage des GAGs.Résultats : L’approche bimodale a permis de mettre en évidence la réponse du CHS SWARM au traitement TH-302 avec une réduction significative de 75% du volume tumoral (p

International audience; Objectifs : La NIRS, comme l'IRMf BOLD, fournit une mesure indirecte de l'activité cérébrale au travers des variations hémodynamiques. Cette technique d'imagerie portable est plus accessible que l'IRMf mais serait moins reproductible, en particulier au niveau individuel [1]. Cette étude évalue la reproductibilité NIRS d'une tâche motrice en test-retest en comparaison avec l'IRMf.

International audience; Aged macular degeneration (AMD) leads to a progressive decline in visual acuity until reaching blindness. It is considered as an irreversible pathology where an early diagnosis remains crucial. However, the lack of ophthalmologists, the permanent increase in elderly people, and their limited mobility involves a delay in AMD diagnosis. In this paper, we propose an automated method for AMD screening. The proposed processing pipeline consists in applying the well-known Radon transform to the macula region in order to model the AMD lesions even with a moderate quality of smartphone-captured fundus images. Thereby, the relevant features are carefully selected, related to the main proprieties of drusens, and then provided to an SVM classifer. The implementation of the method into a smartphone associated to a fundus image capturing device leads to a mobile CAD system that performs higher performance AMD screening. Within this framework and to achieve a real-time implementation, an optimization approach is suggested in order to reduce the processing workload. The evaluation of our method is carried out through the three public STARE, REFUGE, and RFMiD databases. A 4-fold cross-validation approach is used to evaluate the method performance where accuracies of 100%, 95.2%, and 94.3% are respectively obtained with STARE, REFUGE, and RFMiD databases. Comparisons with the state-of-the-art methods in the literature are done. Thereafter, the robustness of the proposed method was evaluated and proved. We note that 100% accuracy was preserved despite the use of degraded quality fundus images as noisy and blurred. Moreover, the propounded method was implemented in S7-Edge and S9 Smartphone devices, where the execution times of 19 and 15 milliseconds were respectively achieved, which proves the AMD real-time detection. Taking advantage of its mobility, cost-efective, detection performance, andreduced execution time, our proposed method seems a good solution for real-time AMD screening on mobile devices.

International audience; Proton computed tomography (pCT) is a promising tomographic imaging modality allowing direct reconstruction of proton relative stopping power (RSP) required for proton therapy dose calculation. In this review article, we aim at highlighting the role of Monte Carlo (MC) simulation in pCT studies. After describing the requirements for performing proton computed tomography and the various pCT scanners actively used in recent research projects, we present an overview of available MC simulation platforms. The use of MC simulations in the scope of investigations of image reconstruction, and for the evaluation of optimal RSP accuracy, precision and spatial resolution omitting detector effects is then described. In the final sections of the review article, we present specific applications of realistic Monte Carlo simulations of an existing pCT scanner prototype, which we describe in detail.

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.

International audience; Cutaneous melanoma is an invasive cancer with a worldwide annual death toll of 57,000 (Arnold \textit{et al.}, JAMA Dermatol 2022). In a metastatic state, surgical interventions are not curative and must be coupled with targeted therapy, or immunotherapy. However, resistance appears almost systematically and late-stage prognosis can remain poor. The complexity to eradicate melanoma stems from its plasticity; these cancer cells continually adapt to the tumor microenvironment, which leads to treatment resistance. Our primary assumption is that therapeutic resistance relies in part on a series of non-genetic transitions including changes in the metabolic states of these cancer cells. The 3D spatial distribution of blood vessels that are sources of nutrition and oxygen that drive this metabolic status is an important variable for understanding zoning aspects of this adaptation process. Using Whole Slide Images (WSI) of melanoma tumors from Patient-Derived Xenograft (PDX) mouse models, we build 3D vascular models to help predict and understand the metabolic states of cancer cells within the tumor. Our 3D reconstruction pipeline was based on PDX tumor samples sectioned over 2mm depth and stained with Hematoxylin and Eosin (H\&E). The pipeline involves three primary steps, including 2D vessel segmentation using Deep Learning, intensity- and affine-based image registration, and 3D reconstruction using interpolation and 3D rendering (allowing for better interaction with biologists, pathologists, and clinicians). The originality of our computer-assisted pipeline is its capability to (a) deal with sparse data (i.e., not all tissue sections were readily available), and (b) adapt to a multitude of WSI-related challenges (e.g., epistemic uncertainty, extended processing times due to WSI scale, etc.). We posit both our 3D reconstruction pipeline, quantitative results of the major stages of the process, and a detailed illustration of the challenges faced, presenting resolutions to improve the pipeline’s efficiency.

Non-invasively estimating brain microstructure that consists of a very large number of neurites, somas, and glial cells is essential for future neuroimaging. Diffusion MRI (dMRI) is a promising technique to probe brain microstructural properties below the spatial resolution of MRI scanners. Due to the structural complexity of brain tissue and the intricate diffusion MRI mechanism, in vivo microstructure estimation is challenging.Existing methods typically use simplified geometries, particularly spheres, and sticks, to model neuronal structures and to obtain analytical expressions of intracellular signals. The validity of the assumptions made by these methods remains undetermined. This thesis aims to facilitate simulationdriven brain microstructure estimation by replacing simplified geometries with realistic neuron geometry models and the analytical intracellular signal expressions with diffusion MRI simulations. Combined with accurate neuron geometry models, numerical dMRI simulations can give accurate intracellular signals and seamlessly incorporate effects arising from, for instance, neurite undulation or water exchange between soma and neurites.Despite these advantages, dMRI simulations have not been widely adopted due to the difficulties in constructing realistic numerical phantoms, the high computational cost of dMRI simulations, and the difficulty in approximating the implicit mappings between dMRI signals and microstructure properties. This thesis addresses the above problems by making four contributions. First, we develop a high-performance opensource neuron mesh generator and make publicly available over a thousand realistic cellular meshes.The neuron mesh generator, swc2mesh, can automatically and robustly convert valuable neuron tracing data into realistic neuron meshes. We have carefully designed the generator to maintain a good balance between mesh quality and size. A neuron mesh database, NeuronSet, which contains 1213 simulation-ready cell meshes and their neuroanatomical measurements, was built using the mesh generator. These meshes served as the basis for further research. Second, we increased the computational efficiency of the numerical matrix formalism method by accelerating the eigendecomposition algorithm and exploiting GPU computing. The speed was increased tenfold. With similar accuracy, the optimized numerical matrix formalism is 20 times faster than the FEM method and 65 times faster than a GPU-based Monte Carlo method. By performing simulations on realistic neuron meshes, we investigated the effect of water exchange between somas and neurites, and the relationship between soma size and signals. We then implemented a new simulation method that provides a Fourier-like representation of the dMRI signals. This method was derived theoretically and implemented numerically. We validated the convergence of the method and showed that the error behavior is consistent with our error analysis. Finally, we propose a simulation-driven supervised learning framework to estimate brain microstructure using diffusion MRI. By exploiting the powerful modeling and computational capabilities that are mentioned above, we have built a synthetic database containing the dMRI signals and microstructure parameters of 1.4 million artificial brain voxels. We have shown that this database can help approximate the underlying mappings of the dMRI signals to volume and surface fractions using artificial neural networks.; L'estimation non invasive de la microstructure du cerveau, qui se compose de nombreux neurites, de somas et de cellules gliales, est essentielle pour l'imagerie cérébrale. L'IRM de diffusion (IRMd) est une technique prometteuse pour sonder les propriétés microstructurelles du cerveau en dessous de la résolution spatiale des scanners IRM. En raison de la complexité structurelle du tissu cérébral et du mécanisme complexe de l'IRM de diffusion, l'estimation de la microstructure in vivo est un défi. Les méthodes existantes utilisent généralement des géométries simplifiées, notamment des sphères et des bâtons, pour modéliser les structures neuronales et obtenir des expressions analytiques des signaux intracellulaires. La validité des hypothèses faites par ces méthodes reste indéterminée. Cette thèse vise à faciliter l'estimation de la microstructure du cerveau par simulation en remplaçant les géométries simplifiées par des modèles réalistes de la géométrie des neurones et les expressions analytiques des signaux intracellulaires par des simulations d'IRM de diffusion. Combinées à des modèles précis de la géométrie des neurones, les simulations numériques d'IRMd peuvent donner des signaux intracellulaires précis et incorporer les effets dus, par exemple, à l'ondulation des neurites ou à l'échange d'eau entre le soma et les neurites.Malgré ces avantages, les simulations d'IRMd n'ont pas été largement adoptées en raison de l'inaccessibilité des fantômes numériques, de la faible efficacité de calcul des simulateurs d'IRMd et de la difficulté d'approximer les mappings implicites entre les signaux d'IRMd et les propriétés de la microstructure. Cette thèse contribue à la résolution des problèmes susmentionnés de la manière suivante : (1) en développant un générateur de maillage de neurones open-source et en rendant accessibles au public plus d'un millier de maillages cellulaires réalistes ; (2) en augmentant d'un facteur dix l'efficacité de calcul de la méthode du formalisme matriciel numérique ; (3) en mettant en œuvre une nouvelle méthode de simulation qui fournit une représentation de type Fourier des signaux IRMd ; (4) en proposant un cadre d'apprentissage supervisé basé sur la simulation pour estimer la microstructure du cerveau par IRM de diffusion.

International audience