Museums increasingly rely on digital content to support visitors' understanding of artworks, yet little is known about how these formats shape the emotional engagement that underlies meaningful art experiences. This research presents an in-situ EEG study on how digital interpretive content modulate engagement during art viewing. Participants experienced three modalities: direct viewing of a Bruegel painting, a 180° immersive interpretive projection, and a regular, display-based interpretive video. Frontal EEG markers of motivational orientation, internal involvement, perceptual drive, and arousal were extracted using eyes-open baselines and Z-normalized contrasts. Results show modality-specific engagement profiles: display-based interpretive video induced high arousal and fast-band activity, immersive projections promoted calm, presence-oriented absorption, and original artworks reflected internally regulated engagement. These findings, relying on lightweight EEG sensing in an operational cultural environment, suggest that digital interpretive content affects engagement style rather than quantity. This paves the way for new multimodal sensing approaches and enables museums to optimize the modalities and content of their interpretive media.

Purpose:Proton therapy is challenged by tumor motion, particularly for lung tumors affected by respiratory-induced motion. Conventional planning strategies compensate for this motion by introducing safety margins or by using robust optimization, increasing irradiation of surrounding healthy tissues. Real-time plan adaptation during delivery represents a promising alternative to mitigate intrafractional motion effects. Materials and Methods:We propose a patient-specific deep Reinforcement Learning (RL)-based control framework for proton pencil beam scanning, formulated as a first step toward a real-time plan adaptation problem under respiratory motion. RL agents are trained independently on 3 patients on their mid-position planning CT to sequentially control beam position and spot delivery. At inference, the learned policy is executed on the respiratory phases of each patient's 4DCT without any online re-training or model parameters adaptation. The agent is provided with 2D observations encoding target geometry, beam position, and prior spot delivery. The approach is evaluated against conventional static gross tumor volume-based (GTV-based) and internal target volume-based (ITV-based) planning strategies. Results:The agents improve target coverage under respiratory motion by exploiting new information in the observations during delivery compared with static gross tumor volume-based plans, with an average gain of 4.54 Gy in D 95 GTV over the whole treatment for Patient 1. Compared with ITV-based plans, the RL-based approach generally reduces dose exposure to organ-at-risk, with an average reduction of 0.42, 4.12, and 3.14 Gy in D mean Lung-GTV for the 3 patients, respectively, and a decrease of 1.41 Gy in D mean Heart for Patient 3, who has the largest motion amplitude. Conclusion:This study should be interpreted as a proof-of-concept and highlights the potential of RL-based control strategies for proton therapy delivery under intrafractional motion. While the current framework relies on static training, it establishes a foundation for future extensions toward fully dynamic and adaptive treatments.

Assisting pathologists in the analysis of histopathological images has high clinical value, as it supports cancer detection and staging. In this context, histology foundation models have recently emerged. Among them, Vision-Language Models (VLMs) provide strong yet imperfect zero-shot predictions. We propose to refine these predictions by adapting Conditional Random Fields (CRFs) to histopathological applications, requiring no additional model training. We present HistoCRF, a CRF-based framework, with a novel definition of the pairwise potential that promotes label diversity and leverages expert annotations. We consider three experiments: without annotations, with expert annotations, and with iterative human-in-the-loop annotations that progressively correct misclassified patches. Experiments on five patch-level classification datasets covering different organs and diseases demonstrate average accuracy gains of 16.0

Background: Accurate forecasting of lung tumor motion is crucial for precise radiotherapy. Deep-learning-based markerless tracking methods have been explored, but extending these approaches to predict future tumor trajectories remains largely unaddressed. We address this by framing markerless lung tumor motion forecasting as a spatio-temporal prediction task using a vision transformer to estimate three-dimensional tumor positions over short horizons. Methods: Digitally reconstructed radiographs (DRRs) generated from four-dimensional computed tomography scans of 12 lung cancer patients were used to train a multi-patient (MP) model. Patient-specific (PS) models trained solely on planning data were compared, and the MP model was further fine-tuned using a small number of patient-specific treatment images under realistic clinical constraints. Models processed sequences of 12 DRRs, with performance evaluated via root mean square error. Results: The results indicate that low-resolution inputs with larger patch sizes outperform higher-resolution configurations by reducing image noise. PS models require extensive data to match MP performance, whereas fine-tuning the MP model with limited patient-specific data achieves comparable or superior forecasting accuracy at a lower cost. Conclusions: These findings demonstrate that Vision Transformers can extend markerless tracking methods to accurate short-term forecasting and highlight fine-tuning as an efficient strategy for personalized prediction.

Multiple sclerosis diagnostic criteria lack optimal specificity, leading to potential misdiagnosis. Advanced magnetic resonance imaging (MRI) biomarkers like the central vein sign, cortical lesions and paramagnetic rim lesions are highly specific to multiple sclerosis and could potentially improve diagnostic accuracy. In this study, we applied machine learning techniques to a retrospective, multicentric dataset of 322 multiple sclerosis/multiple sclerosis-mimic (204/118) and 84 prodromal multiple sclerosis/non-multiple sclerosis (43/41) adult patients, incorporating the central vein sign, cortical lesions and paramagnetic rim lesions. We compared (5 × 2 cross-validation combined F-test) the diagnostic performance of 71 machine learning models, each corresponding to a distinct combination of full-count or simplified biomarker inputs, against the baseline dissemination in space McDonald criteria. The aim was to evaluate the multiple sclerosis diagnostic power of combining these biomarkers in an MRI-only diagnostic framework. 51 of the 71 models significantly outperformed the dissemination in space criterion (P < 0.05), with balanced accuracy improvements up to 13.0% (confidence interval: [+10.5; +17.0]). The best overall model (random forest, using full-count assessments) achieved 95.7% (confidence interval: [93.2; 99.7]) balanced accuracy; the best simplified model (logistic regression, using only simplified assessments) reached 94.7% with no significant difference with the former (P = 0.29). Notably, 12/51 high-performing models used only simplified assessments. To further investigate the models' generalizability, external validation on two out-of-distribution test sets using bootstrapping (1000 resamples) confirmed these results and highlighted a more robust generalization for the best model using solely simplified biomarkers. On the first external test set (n = 37, Verona), the simplified model achieved 97.2% balanced accuracy, while the full-count model reached 93.3% (versus 83.3% for baseline). On the second test set (n = 84, prodromal cases), the simplified model achieved 92.6% (versus 60.1% for baseline) showing competitive performance against the full-count model (93.9%). Both models improved all key performance metrics-balanced accuracy, sensitivity, specificity, precision and F1 score-over the baseline on both test sets (all P < 0.0001). Within a non-invasive MRI-only diagnostic framework, these results show that the incorporation of advanced imaging biomarkers into the multiple sclerosis-MRI diagnostic criteria significantly enhances the diagnostic accuracy-a statement holding true even when using simplified central vein sign, cortical lesions and paramagnetic rim lesions assessments. The study also provides a publicly available online diagnostic tool, facilitating further interaction, validation and clinical support (https://www.msdiagnostictool.org).

The rapid expansion of sensor-based networks introduces major challenges in scalability, adaptability, and knowledge transfer, especially in open environments where new subsystems can dynamically join or leave. In this work, we propose a Holonic Active Distillation architecture within a Holonic Multi-Agent System (HMAS) to address these issues. Our approach integrates Clustered Stream-Based Active Distillation (CSBAD), a framework in which specialized student models collect local data, query pseudo-labels from teacher models, and cluster into groups of similar sensors. Results show that the holonic organization balances local specialization with global generalization, while efficiently adapting to sensor departures and re-integrations. We also analyzed trade-offs among incremental model updates, system reorganization, and scalability limits. Our findings highlight the advantages of holonic learning for multi-sensor systems while identifying key challenges related to model drift and long-term adaptation.

Objective: This study aims to support early diagnosis of Alzheimer's disease and detection of amyloid accumulation by leveraging the microstructural information available in multi-shell diffusion MRI (dMRI) data, using a vision transformer-based deep learning framework. Methods: We present a classification pipeline that employs the Swin Transformer, a hierarchical vision transformer model, on multi-shell dMRI data for the classification of Alzheimer's disease and amyloid presence. Key metrics from DTI and NODDI were extracted and projected onto 2D planes to enable transfer learning with ImageNet-pretrained models. To efficiently adapt the transformer to limited labeled neuroimaging data, we integrated Low-Rank Adaptation. We assessed the framework on diagnostic group prediction (cognitively normal, mild cognitive impairment, Alzheimer's disease dementia) and amyloid status classification. Results: The framework achieved competitive classification results within the scope of multi-shell dMRI-based features, with the best balanced accuracy of 95.2% for distinguishing cognitively normal individuals from those with Alzheimer's disease dementia using NODDI metrics. For amyloid detection, it reached 77.2% balanced accuracy in distinguishing amyloid-positive mild cognitive impairment/Alzheimer's disease dementia subjects from amyloid-negative cognitively normal subjects, and 67.9% for identifying amyloidpositive individuals among cognitively normal subjects. Grad-CAM-based explainability analysis identified clinically relevant brain regions, including the parahippocampal gyrus and hippocampus, as key contributors to model predictions. Conclusion/Significance: This study demonstrates the promise of diffusion MRI and transformer-based architectures for early detection of Alzheimer's disease and amyloid pathology, supporting biomarker-driven diagnostics in data-limited biomedical settings.

While high-capacity AI models have advanced state-of-the-art performance, their practical deployment is often hindered by high inference costs, environmental impact, and a "one-size-fits-all" approach that ignores varying sample complexity. In clinical settings for instance, the waste of computational resources on routine cases is a significant barrier to sustainable AI. In this paper, we introduce the Conformal Adaptive Decision System (CADS), a sequential multi-model algorithm designed to optimize resource allocation by efficiently sampling models based on the estimated data complexity. CADS leverages conformal prediction to quantify image uncertainty at runtime. CADS provides a mathematically grounded framework for balancing the cost-accuracy dilemma that dynamically routes samples through a model cascade, ranging from lightweight "Scout" models to high-capacity "Oracle" architectures. Validated on two datasets, CADS demonstrated superior efficiency and accuracy at a computational cost that can be up to 12 times lower than heavy-model inference. By accurately routing samples based on real-time complexity, CADS ensures high diagnostic reliability while drastically reducing the economic and environmental footprint of AI.

Vision-Language Models for remote sensing have shown promising uses thanks to their extensive pretraining. However, their conventional usage in zero-shot scene classification methods still involves dividing large images into patches and making independent predictions, i.e., inductive inference, thereby limiting their effectiveness by ignoring valuable contextual information. Our approach tackles this issue by utilizing initial predictions based on text prompting and patch affinity relationships from the image encoder to enhance zero-shot capabilities through transductive inference, all without the need for supervision and at a minor computational cost. Experiments on 10 remote sensing datasets with state-of-the-art Vision-Language Models demonstrate significant accuracy improvements over inductive zero-shot classification. Our source code is publicly available on Github: https://github.com/elkhouryk/RS-TransCLIP

Accurate lesion-level segmentation on MRI is critical for multiple sclerosis (MS) diagnosis, prognosis, and disease monitoring. However, current evaluation practices largely rely on semantic segmentation post-processed with connected components (CC), which cannot separate confluent lesions (aggregates of confluent lesion units, CLUs) due to reliance on spatial connectivity. To address this misalignment with clinical needs, we introduce formal definitions of CLUs and associated CLU-aware detection metrics, and include them in an exhaustive instance segmentation evaluation framework. Within this framework, we systematically evaluate CC and post-processing-based Automated Confluent Splitting (ACLS), the only existing methods for lesion instance segmentation in MS. Our analysis reveals that CC consistently underestimates CLU counts, while ACLS tends to oversplit lesions, leading to overestimated lesion counts and reduced precision. To overcome these limitations, we propose ConfLUNet, the first end-to-end instance segmentation framework for MS lesions. ConfLUNet jointly optimizes lesion detection and delineation from a single FLAIR image. Trained on 50 patients, ConfLUNet significantly outperforms CC and ACLS on the held-out test set (n = 13) in instance segmentation (Panoptic Quality: 42.0% vs. 37.5%/36.8%; p = 0.017/0.005) and lesion detection (F1: 67.3% vs. 61.6%/59.9%; p = 0.028/0.013). For CLU detection, ConfLUNet achieves the highest textF1CLU (81.5%), improving recall over CC (+12.5%, p = 0.015) and precision over ACLS (+31.2%, p = 0.003). By combining rigorous definitions, new CLU-aware metrics, a reproducible evaluation framework, and the first dedicated end-to-end model, this work lays the foundation for lesion instance segmentation in MS.

Purpose:To define decade-stratified normative values for vascular parameters obtained from optical coherence tomography angiography (OCTA) in healthy eyes and to evaluate their utility for predicting biological ocular age using artificial intelligence. Methods:This cross-sectional pilot study included 136 rigorously screened healthy subjects aged 10-80 years. Spectral-domain OCT and OCTA scans were acquired using the Optovue Solix platform. Structural and vascular features were extracted from both the macular and optic disc regions. Vessel density (VD) metrics were calculated in the superficial capillary plexus using the ETDRS grid (macula) and Garway-Heath segmentation (peripapillary). Foveal avascular zone (FAZ) area, FAZ circularity, and FD-300 density were also analysed. Disc and RNFL metrics were included. Age-stratified normative values were derived, and a support vector regression (SVR) model was developed to estimate biological ocular age based on structural-only, vascular-only, and combined imaging inputs. Model performance was evaluated using root mean squared error (RMSE) and R² under subject-level grouped cross-validation. Results:Vessel density in the macular and peripapillary regions declined progressively with age, particularly after the fifth decade. FAZ area increased, and circularity decreased with age, while FD-300 density remained relatively stable. The SVR model trained on OCTA-only features showed modest predictive performance (R² = 0.268), while the structural OCT-only model performed poorly (R² = 0.296). Combining structural and vascular features achieved a highly accurate age prediction model (R² = 0.895; RMSE = 5.025 years and MAE of 4.024) under subject-level cross-validation. Conclusion:This pilot study provides decade-stratified normative OCTA metrics and demonstrates that combining OCT and OCTA features significantly enhances AI-based ocular age estimation. These findings offer a promising foundation for early glaucoma risk stratification through biologically meaningful ocular age prediction.

Characterizing domains is essential for models analyzing dynamic environments, as it allows them to adapt to evolving conditions or to hand the task over to backup systems when facing conditions outside their operational domain. Existing solutions typically characterize a domain by solving a regression or classification problem, which limits their applicability as they only provide a limited summarized description of the domain. In this paper, we present a novel approach to domain characterization by characterizing domains as probability distributions. Particularly, we develop a method to predict the likelihood of different weather conditions from images captured by vehicle-mounted cameras by estimating distributions of physical parameters using normalizing flows. To validate our proposed approach, we conduct experiments within the context of autonomous vehicles, focusing on predicting the distribution of weather parameters to characterize the operational domain. This domain is characterized by physical parameters (absolute characterization) and arbitrarily predefined domains (relative characterization). Finally, we evaluate whether a system can safely operate in a target domain by comparing it to multiple source domains where safety has already been established. This approach holds significant potential, as accurate weather prediction and effective domain adaptation are crucial for autonomous systems to adjust to dynamic environmental conditions.

Accurate ground truth estimation in medical screening programs often relies on coalitions of experts and peer second opinions. Algorithms that efficiently aggregate noisy annotations can enhance screening workflows, particularly when data arrive continuously and expert proficiency is initially unknown. However, existing algorithms do not meet the requirements for seamless integration into screening pipelines. We therefore propose an adaptive approach for real-time annotation that (I) supports on-the-fly labeling of incoming data, (II) operates without prior knowledge of medical experts or pre-labeled data, and (III) dynamically queries additional experts based on the latent difficulty of each instance. The method incrementally gathers expert opinions until a confidence threshold is met, providing accurate labels with reduced annotation overhead. We evaluate our approach on three multi-annotator classification datasets across different modalities. Results show that our adaptive querying strategy reduces the number of expert queries by up to 50% while achieving accuracy comparable to a non-adaptive baseline. Our code is available at https://github.com/tbary/MEDICS

As robots increasingly integrate into the workplace, Human-Robot Collaboration (HRC) has become increasingly important. However, most HRC solutions are based on pre-programmed tasks and use fixed safety parameters, which keeps humans out of the loop. To overcome this, HRC solutions that can easily adapt to human preferences during the operation as well as their safety precautions considering the familiarity with robots are necessary. In this paper, we introduce GPTAlly, a novel safety-oriented system for HRC that leverages the emerging capabilities of Large Language Models (LLMs). GPTAlly uses LLMs to 1) infer users’ subjective safety perceptions to modify the parameters of a Safety Index algorithm; 2) decide on subsequent actions when the robot stops to prevent unwanted collisions; and 3) re-shape the robot arm trajectories based on user instructions. We subjectively evaluate the robot’s behavior by comparing the safety perception of GPT-4 to the participants. We also evaluate the accuracy of natural language-based robot programming of decision-making requests. The results show that GPTAlly infers safety perception similarly to humans, and achieves an average of 80% of accuracy in decision-making, with few instances under 50%. Code available at: https://axtiop.github.io/GPTAlly

Objective: Radiotherapy is a primary method for cancer treatment, wherein radiation doses are divided into multiple sessions or fractions to effectively target tumors and minimize damage to surrounding tissues. Methods: In this study, we leverage reinforcement learning (RL) to enhance treatment planning with the aim of improving the adaptability and robustness of RL agents given the inherent inaccuracies in tumor growth models. A 2D simulation model of tumor growth is employed, where tabular RL techniques are used to determine the optimal treatment strategies. We emphasize the significance of tissue damage predictions and incorporate the Lyman NTCP model to assess treatment outcomes, analyzing complications across three simulated body sites: the rectum, head and neck and lung. Results: For all the tumor sites, the RL approach significantly reduces healthy tissue damage by 10.7%, 49.1% and 37.5%, respectively, for rectal, head and neck and lung cancers compared with the baseline treatment. Conclusions: The RL-based approach in radiotherapy not only achieves tumor eradication but also significantly reduces healthy tissue damage compared with traditional treatment methods. This study demonstrates the potential of reinforcement learning to optimize treatment planning in radiotherapy, offering a promising path towards more personalized and effective cancer treatments.

Human-in-the-Loop (HITL) systems are essential in high-stakes, real-world applications where AI must collaborate with human decision-makers. This work investigates how Conformal Prediction (CP) techniques, which provide rigorous coverage guarantees, can enhance the reliability of state-of-the-art human action recognition (HAR) systems built upon Vision-Language Models (VLMs). We demonstrate that CP can significantly reduce the average number of candidate classes without modifying the underlying VLM. However, these reductions often result in distributions with long tails which can hinder their practical utility. To mitigate this, we propose tuning the temperature of the softmax prediction, without using additional calibration data. This work contributes to ongoing efforts for multi-modal human-AI interaction in dynamic real-world environments.

Cytology slides are essential tools in diagnosing and staging cancer, but their analysis is time-consuming and costly. Foundation models have shown great potential to assist in these tasks. In this paper, we explore how existing foundation models can be applied to cytological classification. More particularly, we focus on low-rank adaptation, a parameter-efficient fine-tuning method suited to few-shot learning. We evaluated five foundation models across four cytological classification datasets. Our results demonstrate that fine-tuning the pre-trained backbones with low-rank adaptation significantly improves model performance compared to fine-tuning only the classifier head, achieving state-of-the-art results on both simple and complex classification tasks while requiring fewer data samples. Our source code is available on GitHub https://github.com/mdausort/Cytology-fine-tuning.

Conformal predictions have attracted significant attention in the field of uncertainty quantification, mainly because of their strong marginal coverage guarantees. Full conditional guarantee is not an attainable goal, a well known fact in conformal predictions literature. As a result, several approaches have tried to approximate this behavior by adapting the conformal sets of test-time samples according to their similarity to calibration examples. Although the latter has gained traction and shown impressive performances for regression problems, its application to image classification remains under-explored. We conduct an extensive benchmarking on natural image classification tasks with vision-language models (VLMs), using our open source implementation of a recent localized conformal prediction algorithm. We show that straightforward usage of the cosine similarity between test-time and calibration visual features, an intuitive choice for VLMs, is not sufficient to improve over the non-local baselines. In response, we propose a simple non-linear transformation of the cosine similarities, which conserves marginal coverage guarantees and achieves statistically significant mean set sizes reduction. Code is available at https://github.com/cfuchs2023/lcp-vlm/.

Foundation models provide the adaptability needed in robotics but often require explicit tasks or human verification due to potential unreliability in their responses, complicating human-robot collaboration (HRC). To enhance the reliability of such task-planning systems, we propose 1) an adaptive task-planning system for HRC that reliably performs non-predefined tasks implicitly instructed through HRC, and 2) an integrated system combining multimodal large language model (LLM)-based task planning with multimodal communication of human intention to increase the HRC success rate and comfort. The proposed system integrates GPT-4V for adaptive task planning and comprehension evaluation during HRC with multimodal communication of human intention through speech and deictic gestures. Four pick-and-place tasks of gradually increasing difficulty were used in three experiments, each evaluating a key aspect of the proposed system: task planning, comprehension evaluation, and multimodal communication. The quantitative results show that the proposed system can interpret implicitly instructed tabletop pick-and-place tasks through HRC, providing the next object to pick and the correct position to place it, achieving a mean success rate of 0.80. Additionally, the system can evaluate its comprehension of three of the four tasks with an average precision of 0.87. The qualitative results show that multimodal communication not only significantly enhances the success rate but also the feelings of trust and control, willingness to use again, and sense of collaboration during HRC.