The demand for high-resolution, non-invasive imaging continues to drive innovation in magnetic resonance imaging (MRI), but long acquisition times remain a major practical limitation. Although deep learning-based reconstruction methods have enabled accelerated imaging, their predominant supervised paradigm relies on fully-sampled reference data that are difficult to acquire in practice. Self-supervised learning (SSL) has therefore emerged as a promising alternative, among which splitting methods are a widely used strategy. However, most existing splitting-based methods are empirically designed, and a unified theoretical understanding remains limited. In this work, we introduce UNITS (Unified Theory for Splitting-based self-supervision), a general theoretical framework for splitting-based self-supervised MRI reconstruction. Theoretically, we show that the self-supervised risk can be expressed as a weighted supervised risk. Consequently, self-supervision admits the same pointwise Bayes-optimal predictor as supervised learning. We further relate the training residual to the prediction bias, revealing how different sampling mechanisms affect training behavior. UNITS makes a broad class of existing methods interpretable as special cases within a common framework, and provides a general design space through sampling stochasticity and flexible data utilization. Together, these contributions establish UNITS as a theoretical foundation, a practical paradigm, and a benchmark for interpretable, generalizable, and applicable self-supervised MRI reconstruction.

AI models are increasingly deployed in clinical practice to assist doctors in diagnostic or screening tasks. However, a critical concern arises from the inherent ability of modern AI models to memorise individual examples from their training datasets. Such memorisation could lead to inaccurate predictions when a model is later used on individuals whose historical data was (potentially unknowingly) used for model training or fine-tuning. In this study, we discover evidence for memorisation bias in two large medical imaging datasets: CheXpert (chest radiography) and Kermany-OCT (optical coherence tomography). Our experiments reveal that a small proportion of data-contributing patients ( 0.6% and 1.1% for CheXpert/Kermany-OCT, respectively) exhibit significant changes in their predictions on (future) longitudinal evaluation data when their historical data is included for model training. Strikingly, we find that larger, more diagnostically accurate models exhibit increased memorisation bias: for Kermany-OCT, the number of data-contributing patients affected by memorisation increases substantially (from 1.1% to 7.2% ) when scaling model size from 1.5 to 80 million parameters. Together, our results raise the question whether the future health outcomes of data-contributing patients could be adversely affected by memorisation bias, i.e., predictions which are biased towards their previous health states.

Counterfactual medical image generation aims to modify an existing image to reflect a hypothetical scenario in which certain characteristics of the imaged subject are altered, while keeping their identity fixed. Most existing works repurpose established image editing methods, which do not directly supervise identity preservation. Instead, they assume that identity is implicitly preserved by anchoring generation to the source image. This assumption is rarely tested and may fail in domains where biometric cues are subtle, such as retinal optical coherence tomography (OCT). In this work, we explicitly measure identity preservation for three groups of text-conditioned editing methods - source-anchored, structured-prompt, and paired-training - using referee classifiers, embedding alignment scores, and a blind reader study. We find that all methods produce high-quality OCT images with comparable editing success, yet their identity preservation differs markedly. Source-anchored editing frequently alters the depicted subject, while paired-training preserves it best. We argue that future work on medical counterfactual generation must explicitly measure and report identity preservation alongside image realism and editing success.

We measure how much one extra recurrence is worth to a looped (depth-recurrent) language model, in equivalent unique parameters. From an iso-depth sweep of 116 pretraining runs across recurrence counts r ∈{1, 2, 4, 8} spanning ∼50× in training compute, we fit a joint scaling law L = E + A (N_once + r^φ N_rec)^-α + B D^-β and recover a new recurrence-equivalence exponent φ= 0.46 at R^2 = 0.997. Intuitively, φ tells us whether looping a block r times is equivalent in validation loss to r unique blocks of a non-looped model (full equivalence, φ=1) or to a single block run repeatedly with no capacity gain (φ=0). Our φ= 0.46 sits in between, so each additional recurrence predictably increases validation loss at matched training compute. For example, at r=4 a 410M looped model performs on par with a 580M non-looped model, but pays the training cost of a 1B non-looped one. On a five-axis downstream evaluation, the gap persists on parametric-knowledge tasks and closes on simple open-book tasks, while reasoning tasks are not resolvable at our compute budgets. For any looped LM, our φ converts the design choice of r into a predictable validation-loss cost, and future training recipes and architectures can be compared by how much they raise φ above 0.46.

Neural surrogate models for computational fluid dynamics (CFD) are typically trained as forward operators that map explicit problem specifications, such as geometry and boundary conditions, to solution fields. This ties the model to the conditioning variables seen during training and limits reuse under boundary-condition shifts or local geometry changes. We propose to reformulate steady CFD inference as an inpainting problem: instead of training on explicit boundary conditions, we learn a self-supervised prior over velocity fields and impose boundary constraints only during inference by fixing known regions such as inlet, outlet or unchanged regions from previous simulations. To scale this idea to large 3D meshes, we introduce a local neighbourhood tokeniser that represents high-resolution velocity fields as compact spatial latent tokens and train latent flow-matching and masked-autoencoder models on these tokens. On intracranial aneurysm hemodynamics, our method reconstructs full velocity fields from sparse boundary context, outperforms supervised neural surrogates under boundary-condition and dataset shift and enables local geometry editing by reusing unchanged simulation context. These results suggest that viewing CFD inference as context-conditioned inpainting can turn neural surrogates from task-specific predictors into reusable flow priors.

Large language models (LLMs) are increasingly used to answer health-related questions and support healthcare workflows, yet evidence for their safety still relies heavily on static benchmarks that can rapidly become obsolete or be optimized against. Here we introduce a Dynamic, Automatic and Systematic (DAS) red-teaming audit framework that continuously stress-tests LLMs for health across four safety-critical axes: robustness, privacy, bias and hallucination. Validated against board-certified clinicians, a suite of adversarial agents autonomously mutates health-related test cases to uncover vulnerabilities in real time. Applying DAS to 15 state-of-the-art LLMs revealed a profound gap between high static benchmark performance and low dynamic reliability—the ‘benchmarking gap’. Despite median MedQA accuracy exceeding 80%, 94% of previously correct answers failed under dynamic robustness testing. This brittleness generalized to the realistic, open-ended HealthBench dataset, where top-tier models exhibited failure rates exceeding 70%, suggesting that high scores on established static benchmarks may reflect superficial memorization. We observed similarly high failure rates across other domains: privacy leaks were elicited in 86% of scenarios, cognitive bias priming altered recommendations in 81% of fairness tests and hallucination rates exceeded 74% in widely used models. By converting LLM safety evaluation for health from a static checklist into a living adversarial audit, DAS provides a scalable framework for surfacing latent risks before such systems are deployed in consumer-facing health assistants and broader clinical workflows. Using a suite of adversarial agents designed to test large language models in interactive conversations in a range of safety-related metrics, this study shows strong gaps between static performance and reliability in multi-turn conversations and high failure rates with respect to privacy, fairness and accuracy.

Tool-augmented large language model (LLM) agents can orchestrate specialist classifiers, segmentation models, and visual question-answering modules to interpret chest X-rays. However, these agents still solve each case in isolation: they fail to accumulate experience across cases, correct recurrent reasoning mistakes, or adapt their tool-use behavior without expensive reinforcement learning. While a radiologist naturally improves with every case, current agents remain static. In this work, we propose Evo-MedAgent, a self-evolving memory module that equips a medical agent with the capacity for inter-case learning at test time. Our memory comprises three complementary stores: (1) Retrospective Clinical Episodes that retrieve problem-solving experiences from similar past cases, (2) an Adaptive Procedural Heuristics bank curating priority-tagged diagnostic rules that evolves via reflection, much like a physician refining their internal criteria, and (3) a Tool Reliability Controller that tracks per-tool trustworthiness. On ChestAgentBench, Evo-MedAgent raises multiple-choice question (MCQ) accuracy from 0.68 to 0.79 on GPT-5-mini, and from 0.76 to 0.87 on Gemini-3 Flash. With a strong base model, evolving memory improves performance more effectively than orchestrating external tools on qualitative diagnostic tasks. Because Evo-MedAgent requires no training, its per-case overhead is bounded by one additional retrieval pass and a single reflection call, making it deployable on top of any frozen model.

Reliable recognition and localization of surgical instruments in endoscopic video recordings are foundational for a wide range of applications in computer- and robot-assisted minimally invasive surgery (RAMIS), including surgical training, skill assessment, and autonomous assistance. However, robust performance under real-world conditions remains a significant challenge. Incorporating surgical context - such as the current procedural phase - has emerged as a promising strategy to improve robustness and interpretability. To address these challenges, we organized the Surgical Procedure Phase, Keypoint, and Instrument Recognition (PhaKIR) sub-challenge as part of the Endoscopic Vision (EndoVis) challenge at MICCAI 2024. We introduced a novel, multi-center dataset comprising thirteen full-length laparoscopic cholecystectomy videos collected from three distinct medical institutions, with unified annotations for three interrelated tasks: surgical phase recognition, instrument keypoint estimation, and instrument instance segmentation. Unlike existing datasets, ours enables joint investigation of instrument localization and procedural context within the same data while supporting the integration of temporal information across entire procedures. We report results and findings in accordance with the BIAS guidelines for biomedical image analysis challenges. The PhaKIR sub-challenge advances the field by providing a unique benchmark for developing temporally aware, context-driven methods in RAMIS and offers a high-quality resource to support future research in surgical scene understanding.

Coronary artery stenosis is a leading cause of cardiovascular disease, diagnosed by analyzing the coronary arteries from multiple angiography views. Although numerous deep-learning models have been proposed for stenosis detection from a single angiography view, their performance heavily relies on expensive view-level annotations, which are often not readily available in hospital systems. Moreover, these models fail to capture the temporal dynamics and dependencies among multiple views, which are crucial for clinical diagnosis. To address this, we propose SegmentMIL, a transformer-based multi-view multiple-instance learning framework for patient-level stenosis classification. Trained on a real-world clinical dataset, using patient-level supervision and without any view-level annotations, SegmentMIL jointly predicts the presence of stenosis and localizes the affected anatomical region, distinguishing between the right and left coronary arteries and their respective segments. SegmentMIL obtains high performance on internal and external evaluations and outperforms both view-level models and classical MIL baselines, underscoring its potential as a clinically viable and scalable solution for coronary stenosis diagnosis. Our code is available at https://github.com/NikolaCenic/mil-stenosis.

Multiple myeloma is managed through sequential lines of therapy over years to decades, with each decision depending on cumulative disease history distributed across dozens to hundreds of heterogeneous clinical documents. Whether LLM-based systems can synthesise this evidence at a level approaching expert agreement has not been established. A retrospective evaluation was conducted on longitudinal clinical records of 811 myeloma patients treated at a tertiary centre (2001-2026), covering 44,962 documents and 1,334,677 laboratory values, with external validation on MIMIC-IV. An agentic reasoning system was compared against single-pass retrieval-augmented generation (RAG), iterative RAG, and full-context input on 469 patient-question pairs from 48 templates at three complexity levels. Reference labels came from double annotation by four oncologists with senior haematologist adjudication. Iterative RAG and full-context input converged on a shared ceiling (75.4

Lung cancer risk estimation is gaining increasing importance as more countries introduce population-wide screening programs using low-dose CT (LDCT). As imaging volumes grow, scalable methods that can process entire lung volumes efficiently are essential to tap into the full potential of these large screening datasets. Existing approaches either over-rely on pixel-level annotations, limiting scalability, or analyze the lung in fragments, weakening performance. We present LungEvaty, a fully transformer-based framework for predicting 1-6 year lung cancer risk from a single LDCT scan. The model operates on whole-lung inputs, learning directly from large-scale screening data to capture comprehensive anatomical and pathological cues relevant for malignancy risk. Using only imaging data and no region supervision, LungEvaty matches state-of-the-art performance, refinable by an optional Anatomically Informed Attention Guidance (AIAG) loss that encourages anatomically focused attention. In total, LungEvaty was trained on more than 90,000 CT scans, including over 28,000 for fine-tuning and 6,000 for evaluation. The framework offers a simple, data-efficient, and fully open-source solution that provides an extensible foundation for future research in longitudinal and multimodal lung cancer risk prediction.

To drive progress in science and engineering, large language models (LLMs) must be able to process large amounts of numerical data and solve long calculations efficiently. This is currently only possible through the use of external tools or extensive reasoning chains, either weakening the numerical representations of LLMs or limiting the length of problems they can solve. We show that frontier LLMs require excessive amounts of reasoning tokens to solve even basic calculations, which is exacerbated by their tokenization strategies that split single numbers into multiple tokens. This motivates the need for efficient and effective single-token number encodings. We introduce a set of desiderata for such encodings and show that existing approaches fail to fulfill them. To address these shortcomings, we propose BitTokens, a novel encoding strategy that represents any number as a single token using its IEEE 754 binary floating-point representation. Through extensive experiments we show that our BitTokens allow even small language models to learn algorithms that solve basic arithmetic operations nearly perfectly. This newly gained efficiency could expand the length and complexity of problems language models can solve.

Reusing the early layers of cohort-trained INRs as initialization for new signals has been shown to accelerate and improve signal fitting, yet it remains unclear which layers of the shared encoder learn transferable representations and what those representations encode. We address both questions for two standard backbones, SIREN and Fourier-feature MLPs (FFMLP). First, sweeping the freeze depth across the shared encoder at test time, we find that the optimum coincides with the layer of highest weight stable rank. Moreover, freezing at this depth matches or improves on the standard fine-tuning recipe across all our experiments. Second, identifying which layer transfers does not characterize what that layer encodes. To address this we adopt sparse autoencoders (SAEs), the dominant tool in mechanistic interpretability, and present the first SAE decomposition of INR activations into sparse dictionary atoms. Interestingly, SIREN and FFMLP achieve comparable cohort-fitting quality, but learn qualitatively different dictionaries. Cohort SIREN's atoms are localized, tiling the coordinate plane such that each atom fires in a confined region independent of cohort content. Cohort FFMLP's atoms are image-spanning, tracing the contours of memorized cohort signals. Single-atom ablations confirm causal use of these dictionaries: a single FFMLP atom out of 4096 can drop PSNR by up to 10.6 dB across the image, while SIREN ablations remain confined to where the atom fires. Together, these results give the first mechanistic account of what transfers in cohort-trained INRs and turn their activations into inspectable dictionary atoms. These tools open a path towards characterizing what INRs encode and towards architectures designed for generalization rather than memorization.

Cardiac magnetic resonance (CMR) imaging provides complementary information on cardiac anatomy, function, and tissue characterization across multiple sequences and views. In this work, we investigate foundation model pretraining for 2D CMR and introduce CMRVision, a CMR-specific foundation model trained using DINOv3-style self-supervised learning on a multi-center, multi-sequence cohort of 36 million CMR images. We systematically evaluate architectural and training design choices for domain-specific pretraining. CMRVision is evaluated on two downstream tasks: multi-task segmentation across cine, late gadolinium enhancement (LGE), and mapping sequences, and cine view classification. Our experiments show that CMR-specific pretraining, smaller patch sizes, and patch-level objectives consistently improve downstream performance. Across a multi-task segmentation benchmark, CMRVision achieved the strongest overall performance, outperforming prior natural-image (NI), medical-image, supervised, and CMR foundation model baselines. Improvements were modest but consistent across structures and sequences, with Dice scores ranging from 0.940-0.967 for LV and 0.855-0.905 for myocardium, and reaching 0.929 for RV, 0.920 for LA, and 0.931 for RA. The largest gains were observed for myocardium segmentation in LGE and mapping images. In a zero-shot segmentation task on unseen LGE long-axis views, the model achieved an average Dice score of 0.692, demonstrating cross-view generalization. For cine view classification, CMRVision achieved the highest average accuracy (0.906), compared to prior methods reported in the literature. These results highlight the potential of CMRVision to support robust and generalizable cardiac MRI analysis across multiple sequences and views.

Diagnosing esophageal motility disorders, including dysphagia, pose significant challenges due to the complexity of high-resolution impedance manometry (HRIM) data and variability in clinical interpretation. Thiswork explores the feasibility of a multimodal machine learning (ML)-based classification approach that combines HRIM recordings with patient-specific information and incorporates a graph-based modeling of esophageal physiology. We analyze HRIM recordings with corresponding patient information from 104 patientswith esophageal motility disorders collected atTUMUniversity Hospital. Patient data include demographic, clinical, and symptom information extracted from structured questionnaires and free-text notes using keyword detection and large language model-based processing. HRIM data are represented as spatiotemporal graphs, where nodes correspond to pressure values along the esophagus and edges encode spatial adjacency and impedance dynamics.Agraph neural network (GNN) is applied to learn physiologically meaningful representations,which are fused with patient embeddings for multi-category, multi-class classification of swallow events. The impact of patient features and graphbased modeling is evaluated by ablation studies and comparison to vision-based classifier baselines. The proposed multimodal approach, incorporating patient-specific information, indicates improvements over models that rely solely on HRIM-derived features across all classification categories. Additionally, the graph-based modeling provides gains compared to vision-based baselines. Our experiments systematically assess the complementary contribution of multiple modalities, as well as demonstrate the feasibility of our proposed graph-based approach. Our initial findings demonstrate that integrating patient-level data with graph-based representations of HRIM signals appears to be a promising direction for more accurate classification of esophageal motility disorders. For further validation, future studies should include larger and more representative datasets to confirm these trends and ensure generalizability.

A main promise of looped language models (LMs) is depth-adaptive inference. By iterating a block of shared layers a variable number of times, the model can use less compute for "easy" tokens and more for "hard" ones. However, this adaptivity breaks standard batching: tokens in the same batch now require a different number of loops, so there is no unified forward pass, making efficient inference difficult. Standard inference frameworks like vLLM schedule on the token level and cannot handle this because tokens need to be removed from the batch within the forward pass. Loop-level scheduling has been proposed as a solution, but never implemented end to end. The key challenge is that looped architectures also contain non-looped boundary stages (e.g., token embedding and LM head) that must be scheduled at different frequencies than the loop. We introduce continuous depth batching (CDB), which schedules at the granularity of individual loop iterations. CDB handles boundary stages and loop steps in separate priority queues, makes exit decisions one step ahead, and overlaps all scheduling work with GPU computation. On Ouro 1.4B and Huginn 3.5B, CDB can realize up to 99% of the theoretical maximum speed-up from adaptive-depth, translating to 1.5-1.9× higher offline throughput and 45-90% lower normalized latency under dynamic serving load.

Whole-slide images (WSIs) are fundamental for computational pathology, where accurate lesion segmentation is critical for clinical decision making. Existing methods partition WSIs into discrete patches, disrupting spatial continuity and treating multi-resolution views as independent samples, which leads to spatially fragmented segmentation and reduced robustness to resolution variations. To address the issues, we propose WSI-INR, a novel patch-free framework based on Implicit Neural Representations (INRs). WSI-INR models the WSI as a continuous implicit function mapping spatial coordinates directly to tissue semantics features, outputting segmentation results while preserving intrinsic spatial information across the entire slide. In the WSI-INR, we incorporate multi-resolution hash grid encoding to regard different resolution levels as varying sampling densities of the same continuous tissue, achieving a consistent feature representation across resolutions. In addition, by jointly training a shared INR decoder, WSI-INR can capture general priors across different cases. Experimental results showed that WSI-INR maintains robust segmentation performance across resolutions; at Base/4, our resolution-specific optimization improves Dice score by +26.11

AI-based image reconstruction models are increasingly deployed in clinical workflows to improve image quality from noisy data, such as low-dose X-rays or accelerated MRI scans. However, these models are typically evaluated using pixel-level metrics like PSNR, leaving their impact on downstream diagnostic performance and fairness unclear. We introduce a scalable evaluation framework that applies reconstruction and diagnostic AI models in tandem, which we apply to two tasks (classification, segmentation), three reconstruction approaches (U-Net, GAN, diffusion), and two data types (X-ray, MRI) to assess the potential downstream implications of reconstruction. We find that conventional reconstruction metrics poorly track task performance, where diagnostic accuracy remains largely stable even as reconstruction PSNR declines with increasing image noise. Fairness metrics exhibit greater variability, with reconstruction sometimes amplifying demographic biases, particularly regarding patient sex. However, the overall magnitude of this additional bias is modest compared to the inherent biases already present in diagnostic models. To explore potential bias mitigation, we adapt two strategies from classification literature to the reconstruction setting, but observe limited efficacy. Overall, our findings emphasize the importance of holistic performance and fairness assessments throughout the entire medical imaging workflow, especially as generative reconstruction models are increasingly deployed.

Background. Low-dose computed tomography (LDCT) screening reduces lung cancer mortality fundamentally through patient stratification, determining management intensity and screening intervals. Clinical guidelines such as Lung-RADS and NELSON make this assignment from nodule features, which can be measured by a radiologist or by a deep learning feature extractor. Emerging end-to-end deep learning approaches predict multi-year lung cancer (continuous) risk from a single LDCT, independent of nodule presence. These models are evaluated almost entirely by aggregate discrimination — AUC and the concordance index — which summarize ranking but do not localize where across the risk distribution a model is useful. They are not assessed as the actionable stratification tools they are meant to be. Methods. We propose an evaluation framework for end-to-end models that is aligned with clinical deployment and rooted in lung screening standards. It retains high-level ranking metrics (AUC, PR-AUC, concordance index) for comparability, and adds operating-point evaluation anchored to Lung-RADS: cancer capture from Lorenz curves, clinical utility from decision-curve analysis (DCA), calibration assessed both overall and within Lung-RADS categories. Because every operating point is a Lung-RADS operating point, the framework compares models with one another and with clinical standards such as Lung-RADS or NELSON. We applied it to three contemporary models that predict 6-year risk from a single LDCT on the National Lung Screening Trial (NLST) test cohort (6,223 scans / 2,199 patients), with retrospectively assigned Lung-RADS as the clinical standard and the non-imaging PLCOm2012 model as a baseline. Findings. The models have similar AUCs with overlapping confidence intervals (year-1 AUC 0.91–0.95). Calibration was good in aggregate but varied by category: the models over-predicted in the low-risk Lung-RADS tiers, were well calibrated at intermediate risk, and under-predicted in the high-risk tiers where workup is decided (11% predicted against 22% observed at the highest-risk tier). On cumulative-gains curves and decision-curve analysis, the stronger end-to-end models achieved at least on-par or higher cancer capture and greater net benefit than Lung-RADS, and the advantage was largest on baseline scans, where Lung-RADS has no prior CT for comparison, exceeding it by more than 25 percentage points in capture. Furthermore, the continuous scores were also effective at the extremes the categories cannot reach: they isolated very-high-risk patients at flag rates below the smallest Lung-RADS category, and recovered and ranked higher-risk patients placed in the benign Lung-RADS tiers, where the top score decile held about 70% of the cancers hidden in those tiers (6.7–7.4x enrichment). Implications. Evaluating end-to-end models as stratification tools, rather than by global discrimination, assesses them on the task they are designed for, and places them in the same frame as the clinical standards they would support: Lung-RADS, and feature-based deep learning models that classify by its rules. In that frame, end-to-end deep learning and Lung-RADS are not best understood as competitors. They derive risk differently and have complementary strengths, and our results indicate they are more useful combined than ranked against each other.