Proton Magnetic Resonance Spectroscopy (MRS 1H) at ultra-high field (7T) enables improved metabolic characterization of brain tissues but remains limited by challenges in spectral fitting accuracy and quantification reliability. In this study, we aimed to enhance metabolite quantification using Physics-Informed Neural Network (PINN) and to evaluate the impact of improved quantification on classification model performance. We developed a new PINN-based spectral fitting framework, named PINN-COLLOC, incorporating collocation points to locally constrain the learning process in complex spectral regions. The proposed method was evaluated against LCModel, the current clinical reference and our previous PINN implementation (PINN-IG). Results demonstrated significant improvements in spectral fitting accuracy for several metabolites, particularly low-concentration or highly overlapping resonances such as lactate, glutamine, lipids and phosphocholine + glycerophosphocholine. Ablation studies confirmed the importance of physical constraints and collocation points in improving model robustness. Beyond spectral fitting, we assessed the impact of metabolite quantification on artificial intelligence classification tasks. A classification model trained using metabolite concentrations derived from PINN-based quantification showed improved diagnostic performance compared to LCModel-based quantification, highlighting the potential clinical relevance of improved spectral fitting. Importantly, our PINN models remove dependency on conventional basis sets, reducing limitations related to their completeness and standardization. However, current implementations still rely on post-processing steps and external parameter extraction tools. Future work will focus on developing a fully standardized framework (PINN-FID) operating directly on time-domain free induction decay signals to eliminate post-processing variability and improve reproducibility across clinical centers. The code is available at https://github.com/lbonnin1/MRS-PINN-COLLOC

Gadolinium-based contrast agents (GBCAs) are widely used in magnetic resonance imaging (MRI) to enhance lesion detection and characterisation, particularly in the field of neuro-oncology. Nevertheless, concerns regarding gadolinium retention and accumulation in brain and body tissues, most notably for diseases that require close monitoring and frequent GBCA injection, have led to the need for strategies to reduce dosage. In this study, a deep learning framework is proposed for the virtual contrast enhancement of full-dose post-contrast T1-weighted MRI images from corresponding low-dose acquisitions. The contribution of the presented model is its utilisation of longitudinal information, which is achieved by incorporating a prior full-dose MRI examination from the same patient. A comparative evaluation against a non-longitudinal single session model demonstrated that the longitudinal approach significantly improves image quality across multiple reconstruction metrics. Furthermore, experiments with varying simulated contrast doses confirmed the robustness of the proposed method. These results emphasize the potential of integrating prior imaging history into deep learning-based virtual contrast enhancement pipelines to reduce GBCA usage without compromising diagnostic utility, thus paving the way for safer, more sustainable longitudinal monitoring in clinical MRI practice.

AI-driven image-to-image synthesis is rapidly advancing, with growing applications in medical imaging. Multi-modal image analysis plays a crucial role in optimizing examination quality, yet acquiring multiple imaging modalities in clinical settings remains resource-intensive and time-consuming, especially for 3D imaging. To address this challenge, we propose a novel image-to-image translation model based on Brownian Bridge Diffusion Models (BBDM), which synthesizes magnetic resonance imaging (MRI) sequences from 2D axial slices. Our approach integrates a variational encoder-guided diffusion mechanism, leveraging probabilistic image distributions to enhance synthesis quality. Evaluated on the BraTS 2021 dataset, our Probabilistic-BBDM (Prob-BBDM) achieves superior performance across multiple translation tasks, reaching up to 88.46% SSIM and 26.09 dB PSNR, with consistent improvements over baselines. Notably, our diffusion process requires only 4 steps, making it computationally efficient while maintaining high-quality synthesis. To further validate generalizability, we test Prob-BBDM on an external third-party dataset, demonstrating consistent performance across domains. Additionally, we assess the clinical utility of the synthesized slices by using them as input to a pre-trained segmentation model. Tumor segmentation yields a Dice score of 88.71% and an HD95 of 3.49mm, confirming that the synthesized slices preserve critical diagnostic information. These results highlight the potential of Prob-BBDM for high-quality, efficient, and generalizable MRI synthesis, offering a promising step toward improved medical image translation. https://gitlab.xlim.fr/mvalls/Prob-BBDM.

Physics-Informed Neural Networks (PINNs) constitute a new type of neural architectures that incorporate governing physical laws and principles directly into the training process. This hybrid data-physics framework has attracted significant attention in domains where physical principles guide the interpretation and modeling of data, notably in neuroscience and brain imaging analysis (BIA). The cost of misinterpretation and misdiagnosis in this field is high, especially in the case of neurodegenerative disorders such as Alzheimer's disease (AD). PINNs offer a promising solution to persistent challenges, including the scarcity of labeled data, high noise levels, and the complexity of inverse problems. This paper reviews 9 recent studies, published between 2020 and 2025, that apply PINN-based models to brain imaging in AD. We focus on their integration into core tasks, such as image reconstruction, classification, predictive modeling, and biomarker analysis. We dive in-depth into the role of PINNs in diagnosing and monitoring AD, highlighting how they can support the detection of structural, functional, and metabolic brain alterations, contributing to earlier and more accurate diagnoses.

Anna is one of the 1.8 million people worldwide with multiple sclerosis who live with the uncertainty of disease progression every day [1]. Traditional Magnetic Resonance Imaging scans every six months reveal brain lesions but can't predict how the disease will progress [2]. A new technology, Magnetic Resonance Spectroscopy (MRS), shows promise in predicting disease progression by revealing cerebral metabolism and neurophysiological changes [3]. However, current MRS measurement methods vary between medical centers, affecting reliability [[4], [5], [6]]. Standardizing these measurements using Physics-Informed Neural Networks (PINNs), which are more reliable than traditional neural networks because they are based on the physics of spectra, could ensure accurate, comparable results worldwide [[7], [8], [9]]. This would reassure doctors and patients like Anna, and potentially improve their quality of life by enabling earlier and more precise treatment.

DSC-MRI perfusion is a medical imaging technique for diagnosing and prognosing brain tumors and strokes. Its analysis relies on mathematical deconvolution, but noise or motion artifacts in a clinical environment can disrupt this process, leading to incorrect estimates of perfusion parameters. Although deep learning approaches have shown promising results, their calibrations typically rely on third-party deconvolution algorithms to generate reference outputs and are bound to reproduce their limitations. To address this problem, we propose a physics-informed autoencoder that uses an analytical model to decode the perfusion parameters and guide the learning of the encoding network. This autoencoder is trained in a self-supervised fashion without any third-party software and its performance is evaluated on a database with glioma patients. Our method shows reliable results for glioma grading in accordance with other well-known deconvolution algorithms despite a lower computation time. It also achieved competitive performance even in the presence of high noise, which is critical in a medical environment.

Many approaches have been taken to the task of fitting Proton Magnetic Resonance Spectroscopy ( ^1 H-MRS) data, but the challenge remains unsolved. Traditional neural networks for example perform poorly in the fitting of low quantity metabolites and require large datasets. Recently, Physics Informed Neural Networks (PINN) has emmerged as a promising solution for fitting ^1 H-MRS data by integrating physical laws with artificial intelligence. This paper proposes a PINN framework for fitting ^1 H-MRS 7T data that outperforms current methods with a speed of 0.030 ms for the fitting, robust to pathological variations, providing accurate and rapid fitting of brain metabolites relevant in the diagnosis of the central nervous system diseases and disorders. Source code is available at https://github.com/lbonnin1/MRS-PINN-GUI.git .

Traditional neural networks have shown promising performance in fitting proton Magnetic Resonance Spectroscopy (1H-MRS). However, they have difficulty adapting to low abundance metabolites and have limitations that make them difficult to use in clinical applications: the training database needs to be large and they do not take into account the physical reality of 1H-MRS. Recently, a new neural network method has emerged, the Physics Informed Neural Network (PINN), which links neural networks to physical laws. It provides a promising solution for fitting 1H-MRS data. In this paper, we propose a PINN framework for fitting 1H-MRS 7T data. The proposed method consists of an end-to-end one-dimensional CNN encoder-decoder model integrating a new method of discriminative spectral parameter extraction and weight adjustment guided by the physical reality of 1H-MRS. The data used in this article are collected in our center, which contain 90 1H-MRS samples, healthy and pathological, from multilobe white and grey matter. The proposed framework gives similar and better results than the current scientific literature with a speed of 0.030 ms for the 1H-MRS 7T fitting task. Thus, our model provide accurate and fast fitting of brain metabolites in clinical contexts that can discriminate pathologies.

Nowadays, deep prediction models, especially graph neural networks, have a majorplace in critical applications. In such context, those models need to be highlyinterpretable or being explainable by humans, and at the societal scope, this understandingmay also be feasible for humans that do not have a strong prior knowledgein models and contexts that need to be explained. In the literature, explainingis a human knowledge transfer process regarding a phenomenon between an explainerand an explainee. We propose EiX-GNN (Eigencentrality eXplainer forGraph Neural Networks) a new powerful method for explaining graph neural networksthat encodes computationally this social explainer-to-explainee dependenceunderlying in the explanation process. To handle this dependency, we introducethe notion of explainee concept assimibility which allows explainer to adapt itsexplanation to explainee background or expectation. We lead a qualitative studyto illustrate our explainee concept assimibility notion on real-world data as wellas a qualitative study that compares, according to objective metrics established inthe literature, fairness and compactness of our method with respect to performingstate-of-the-art methods. It turns out that our method achieves strong results inboth aspects.

Face expressions understanding is a key to have a better understanding of the human nature. In this contribution we propose an end-to-end pipeline that takes color images as inputs and produces a semantic graph that encodes numerically what are facial emotions. This approach leverages low-level geometric details as face representation which are numerical representations of facial muscle activation patterns to build this emotional understanding. It shows that our method recovers social expectations of what characterize facial emotions.

Currently, the default brain mode known in the scientific literature uses hemodynamic or electrical data to assess the activity and connection of different brain structures in the basal state. But these data are measured at random times of the day. Conversely, research has shown that every structure in the human body has its own physiological dynamics throughout the day. We therefore designed and carried out a study to assess, this time on the basis of metabolic data, the diurnal brain dynamics in nine different brain structures. The aim was also to establish an initial overview of the brain's metabolic connections. Our results are encouraging and in line with the scientific literature. We were able to identify the diurnal metabolic dynamics, as a function of circadian rhythm, of the nine brain zones in the basal state, as well as certain correlations between them. Furthermore, we observed that in two correlated zones, metabolites follow the same dynamics but not in the same quantities. Finally, the diurnal metabolic dynamics of lactate are surprising. None of the nine zones correlate with each other. In addition, in the same brain structure, its diurnal metabolic dynamics are also independent of other metabolites.

In the pneumatic industry, quality control is an essential step in assessing tire compliance. Artificial neural networks are increasingly used to accomplish this task. Their training requires a large number of images of the controlled products. However, at the launch production of a new tire, the lack of images causes a performance loss for the network. To solve this problem, we propose to translate perfect tires computer-based images into ad hoc manufacturing context-realistic ones as pre-manufacturing step to improve robustness and ensure production quality. The challenging work is to extract features in real images and apply them to computer-based images while maintaining the original geometry. In the paper, we propose Prefab-GEN, a novel architecture based on Cycle-GAN. In the generator part, an Inception U-Net architecture is developed to enforce geometrical structure conversion and extract more detailed features. The qualitative and quantitative evaluation on tire dataset shows improvements compared with state-of-art.

L'Imagerie par Résonance Magnétique (IRM) permet de construire, de façon non-invasive, des représentations visuelles d'une précision inégalable pour différencier des tissus. Une séquence IRM, si précise soit elle, ne peut mettre en évidence la totalité des structures anatomiques et fonctionnelles connues. La production de différentes séquences T1, T2, FLAIR, diffusion, ou encore T2* est nécessaire mais très couteuse en temps machine. En outre, une injection de produit de contraste gadoliné peut être requise, affectant le principe d'une médicine non-invasive. Si la théorie de l'information nous apprend qu'une information robuste est une information redondante, elle nous révèle également qu'il existe des espaces latents de représentation des données éliminant toute redondance. Ce principe de compression, bien connu lors du stockage numérique des données, peut théoriquement être mis à profit dès leur acquisition. Nous proposons un dispositif d'Intelligence Artificielle (IA) permettant de limiter le coût d'examen et l'injection de produit de contraste en générant une partie des contenus IRM de façon synthétique à partir d'un ensemble réduit de séquences. Celui-ci part du principe que les séquences non-acquises présentent un caractère redondant, et se base sur un paradigme récent de synthèse d'images en IA : les réseaux adverses génératifs.

Magnetic resonance imaging (MRI) provides detailed anatomical information critical for radiologists in assessing and diagnosing. Moreover, complementary information obtained from multiple contrasts MRI can improve the clarity of internal structures, especially in detecting unusual objects. However, acquiring contrast-enhanced MRI is usually time-consuming, expensive, or requires contrast agent injection. Medical image synthesis has been demonstrated as an effective alternative. Within the scope of project, we aim to provide a non-invasive method to synthesize contrast-enhanced MRI from a given MRI modality. We present different generative frameworks to learn the mappings between T1-weighted and contrasted Tl MRI. Frameworks jointly exploit different features between cross-modalities to resolve the challenging complexity in synthesis. Methods are trained on a multimodal brain MRI dataset of different contrast samples. Quantitative assessments were conducted by computing peak signal-to-noise ratio and structural similarity index measurements. Results on synthesized output are clear with a low distortion, showing the potential of study in practice.

1) Objectives: Machine learning for binary glioma grading have been extensively used on anatomical MRI, especially using the BraTS dataset. The relevance of radiomic criteria based on multimodal imaging, including diffusion, perfusion and spectroscopy data is to be explored, as multimodal datasets are scarce, and there is no common benchmark for performance comparison. 2) Material and methods: Poitiers University Hospital provides 123 multimodal patient data. We computed 124 features and let a recursive feature elimination algorithm (RFE) yield a relevant, reduced subset of features. We trained a SVM classifier on this subset. We proposed a method to adapt the BraTS dataset to allow performance comparison with the literature. We got a performance reference point by training on anatomical data only, and showed improvements when multimodalities were added. We explored the feature relevance through the RFE subset. The RFE subset is not constant and induce variability in the performances. To smooth the variability, we applied the RFE algorithm 100 times and incremented the selected features, resulting in a global feature ranking. We also show the best classifier reached on these 100 trainings and its feature subset. 3) Results: The best classifier reached 86.5% accuracy, with a mean accuracy on 100 trainings of 78.6%. The rankings shows that anatomical and perfusion sequences are the most relevant for glioma grading, especially T1 post-gadolinium, cerebral blood volume and flow. Intensity and texture features are frequently selected, while anisotropic diffusion coefficient, time to peak and mean time transit mappings seem irrelevant. 4) Conclusion: Multimodal radiomics improve the classification and are consistent with the radiological analysis.

Deep learning has shown record-shattering performance in multiple medical tasks. However, data quantity and quality are crucial requirements. As a matter of fact, data is one of the most challenging issues while deploying deep learning models for different tasks. One of the main challenges is the institutions' privacy protocols, in particular in the medical field. Indeed, the metadata is usually excluded from the database provided. Many invisible features in images can help tracing anonymized data. We propose to use deep learning to exclude these traces. This article focuses on Magnetic resonance imaging (MRI) and one of the most important features, the equipment used for acquisition. First, we aim to produce an algorithm able to perform well distinguishing multiple MRI equipment from different brands. To this end, we employ a convolution neural network architecture to work on this medical image classification task. The second part of this paper is dedicated to reconstructing the input MRI using a simple auto-encoder. The latter step is to use the auto-encoder in order to mislead the classifier classifying the MRI equipment.

Glioma is one of the most important central nervous system tumors, ranked 15th in the most common cancer for men and women. Magnetic Resonance Imaging (MRI) represents a common tool for medical experts to the diagnosis of glioma. A set of multi-sequences from an MRI is selected according to the severity of the pathology. Our proposed approach aims moreto create a computer-aided system that is capable of helping morethe expert diagnose the brain gliomas. moreWe propose a supervised learning regime based on a convolutional neural network based framework and transfer learning techniques. Our research morefocuses on the performance of different pre-trained deep learning models with respect to different MRI sequences. We highlight the best combinations of such model-MRI sequence couple for our specific task of classifying healthy brain against brain with glioma. moreWe also propose to visually analyze the extracted deep features for studying the existing relation of the MRI sequences and models. This interpretability analysis gives some hints for medical expert to understand the diagnosis made by the models. Our study is based on the well-known BraTS datasets including multi-sequence images and expert diagnosis.

Purpose: To evaluate the usefulness of computed tomography (CT) texture descriptors integrated with machine-learning (ML) models in the identification of clear cell renal cell carcinoma (ccRCC) and for the first time papillary renal cell carcinoma (pRCC) tumor nuclear grades [World Health Organization (WHO)/International Society of Urologic Pathologists (ISUP) 1, 2, 3, and 4].Approach: A total of 143 ccRCC and 21 pRCC patients were analyzed in this study. Texture features were extracted from late arterial phase CT images. A complete separation of training/validation and testing subsets from the beginning to the end of the pipeline was adopted. Feature dimension was reduced by collinearity analysis and Gini impurity-based feature selection. The synthetic minority over-sampling technique was employed for imbalanced datasets. The ML classifiers were logistic regression, SVM, RF, multi-layer perceptron, and K-NN. The differentiation between low grades/ high grades, grade 1/grade 2, grade 3/grade 4, and between all grades was assessed for ccRCC and pRCC datasets. The classification performance was assessed and compared by certain metrics.Results: Textures-based classifiers were able to efficiently identify ccRCC and pRCC grades. An accuracy and area under the characteristic operating curve (AUC) up to 91%/0.9, 91%/0.9, 90%/0.9, and 88%/1 were reached when discriminating ccRCC low grades/ high grades, grade 1/grade 2, grade 3/grade 4, and all grades, respectively. An accuracy and AUC up to 96%/1, 81%/0.8, 86%/0.9, and 88%/0.9 were found when differentiating pRCC low grades/ high grades, grade 1/grade 2, grade 3/grade 4, and all grades, respectively.Conclusion: CT texture-based ML models can be used to assist radiologist in predicting the WHO/ISUP grade of ccRCC and pRCC pre-operatively.