Flow matching-based generative models have been integrated into the plug-and-play image restoration framework, and the resulting plug-and-play flow matching (PnP-Flow) model has achieved some remarkable empirical success for image restoration. However, the theoretical understanding of PnP-Flow lags its empirical success. In this paper, we derive a continuous limit for PnP-Flow, resulting in a stochastic differential equation (SDE) surrogate model of PnP-Flow. The SDE model provides two particular insights to improve PnP-Flow for image restoration: (1) It enables us to quantify the error for image restoration, informing us to improve step scheduling and regularize the Lipschitz constant of the neural network-parameterized vector field for error reduction. (2) It informs us to accelerate off-the-shelf PnP-Flow models via extrapolation, resulting in a rescaled version of the proposed SDE model. We validate the efficacy of the SDE-informed improved PnP-Flow using several benchmark tasks, including image denoising, deblurring, super-resolution, and inpainting. Numerical results show that our method significantly outperforms the baseline PnP-Flow and other state-of-the-art approaches, achieving superior performance across evaluation metrics. 

Legg-Calvé-Perthes Disease (LCPD) often results in a permanent residual hip deformity, with long-term impacts on hip function. In the clinic, decisions on how to manage LCPD deformity are based on 2D radiographic measures such as the Stulberg grades. It is unclear whether these measures accurately represent complex, patient-specific 3D deformity, making it difficult to determine the best treatment approach. Statistical shape modeling (SSM) provides an objective and compact description of variability, but there is currently no 3D SSM that focuses solely on describing shape variability in LCPD. In this study, we have constructed an SSM of femurs with LCPD deformity, which will provide a foundation for future research into how deformity is measured in the clinic.

Clinical magnetic resonance (MR) images (N=13 hips, 11 patients, 3 F/8 M) of affected hips were obtained in patients with stage IV LCPD (age range: 6–12 years, estimated Stulberg grades: II-V). MR volumes were resampled isotropically to the smallest voxel dimension (.47 – .63 mm), and two raters manually segmented the proximal femurs. ShapeWorks (SCI Institute, University of Utah, Salt Lake City, UT) was used to produce an SSM with 512 particles using an incremental optimization routine: corresponding particles were initially distributed and optimized on a subset of the five most similar femurs. The mean particle distribution of the initial model was then used to initialize the remaining eight shapes, followed by an additional optimization of the particle distribution. Differences in pose and scale between shapes were removed with generalized Procrustes analysis. Modes of shape variation were quantified using principal component analysis.

The first four shape modes described 87.5% of the population variability, and qualitatively captured clinically relevant variables such as femoral head width and articular-trochanteric distance (see Figure 1). A significant association (p=0.04) was observed between shape mode 4 and asphericity of the femoral head (rho=0.58). A leave-one-out cross-validation of the model demonstrated good generalizability to unseen shapes, with a mean point-to-point reconstruction error of < 1 mm if > 4 modes were included.

This SSM provides a continuous, accurate, and compact representation of 3D shape variation in LCPD. Limitations to this model include a small sample size, and thus the inability to control for age. In addition to expanding this model, future work will incorporate the acetabular side to describe joint congruence, and associate shape modes with radiographic parameters.

Increasing the fidelity of analyses of nuclear material is an important goal of technical nuclear forensics R&D. Many well-proven and mature analysis tools exist; some give visual representations of materials, while others offer clues to the chemical composition. Two such tools are Scanning Electron Microscopy (SEM) and X-ray diffraction (XRD). We present a machine learning framework for fusing these distinct modalities—themselves well-studied tools of nuclear forensics—to increase their specificity together. Early model fusion allows the use of correlations between data sources. We significantly improve accuracy over any single model alone, even over baseline methods that incorporate ground-truth information about the material. If XRD data is limited, we show that generating synthetic patterns from crystallographic information is viable for increasing available data. Comparing models trained on real and synthetic XRD pattern data, we conclude that our early fusion method trained on real XRD pattern data can utilize processing route information previously hidden in the XRD pattern, outperforming late fusion methods and models trained on synthetic data alone.

Motivation:
Hemodynamic analysis is crucial for diagnosing and predicting cardiovascular diseases. However, methods relying on fluid flow simulations or blood flow imaging are complex, time-consuming, and require specialized expertise, limiting their clinical use.

Goal:
This research aims to automate the enhancement of blood flow images, providing clinicians with a fast, accurate tool for hemodynamic analysis without requiring advanced expertise.

Objectives:
A software tool based on physics-constrained neural networks was developed to enable clinicians to easily select and process regions of interest (ROIs) in time-resolved three-dimensional phase contrast magnetic resonance imaging (4D-Flow MRI) blood flow images for quick, accurate analysis.

Methods:
The Input Parameterized Physics-Informed Neural Network (IP-PINN) was introduced to improve the spatio-temporal resolution of 4D-Flow MRI. IP-PINN mitigates noise, velocity aliasing, and phase errors. A convolutional neural network processes ROI data into latent vectors, which are then used to predict velocity, pressure, and spin density via a multi-layer perceptron. The method is trained with synthetic blood flow data using an innovative loss function that addresses noise and artifacts.

Results:
IP-PINN successfully enhanced image resolution, reducing noise and artifacts when tested on synthetic 4D-Flow MRI data derived from blood flow simulations of intracranial aneurysms. For data with 20 decibels (dB) signal-to-noise ratio, results closely matched the ground truth with less than 5.5% relative error. Processing took under two minutes. The method also has the potential to reduce data acquisition time by 25%.

Conclusions:
IP-PINN could significantly enhance the clinical use of 4D-Flow MRI for personalized hemodynamic analysis in cardiovascular diseases.

Spatial transformations that capture population-level morphological statistics are critical for medical image analysis. Commonly used smoothness regularizers for image registration fail to integrate population statistics, leading to anatomically inconsistent transformations. Inverse consistency regularizers promote geometric consistency but lack population morphometrics integration. Regularizers that constrain deformation to low-dimensional manifold methods address this. However, they prioritize reconstruction over interpretability and neglect diffeomorphic properties, such as group composition and inverse consistency. We introduce MORPH-LER, a Log-Euclidean regularization framework for population-aware unsupervised image registration. MORPH-LER learns population morphometrics from spatial transformations to guide and regularize registration networks, ensuring anatomically plausible deformations. It features a bottleneck autoencoder that computes the principal logarithm of deformation fields via iterative square-root predictions. It creates a linearized latent space that respects diffeomorphic properties and enforces inverse consistency. By integrating a registration network with a diffeomorphic autoencoder, MORPH-LER produces smooth, meaningful deformation fields. The framework offers two main contributions: (1) a data-driven regularization strategy that incorporates population-level anatomical statistics to enhance transformation validity and (2) a linearized latent space that enables compact and interpretable deformation fields for efficient population morphometrics analysis. We validate MORPH-LER across two families of deep learning-based registration networks, demonstrating its ability to produce anatomically accurate, computationally efficient, and statistically meaningful transformations on the OASIS-1 brain imaging dataset.

Hematoxylin and eosin (H&E) staining is a gold standard for microscopic diagnosis in pathology. However, H&E staining does not capture all the diagnostic information that may be needed. To obtain additional molecular information, immunohistochemical (IHC) stains highlight proteins that mark specific cell types, such as CD3 for T-cells or CK8/18 for epithelial cells. While IHC stains are vital for prognosis and treatment guidance, they are typically only available at specialized centers and time consuming to acquire, leading to treatment delays for patients. Virtual staining, enabled by deep learning-based image translation models, provides a promising alternative by computationally generating IHC stains from H&E stained images. Although many GAN and diffusion based image to image (I2I) translation methods have been used for virtual staining, these models treat image patches as independent data points, which results in increased and more diverse data requirements for effective generation. We present ImplicitStainer, a novel approach that leverages local implicit functions to improve image translation, specifically virtual staining performance, by focusing on pixel-level predictions. This method enhances robustness to variations in dataset sizes, delivering high-quality results even with limited data. We validate our approach on two datasets using a comprehensive set of metrics and benchmark it against over fifteen state-of-the-art GAN- and diffusion based models. Full Code and models trained will be released publicly via Github upon acceptance.

Obtaining large-scale medical data, annotated or unannotated, is challenging due to stringent privacy regulations and data protection policies. In addition, annotating medical images requires that domain experts manually delineate anatomical structures, making the process both time-consuming and costly. As a result, semi-supervised methods have gained popularity for reducing annotation costs. However, the performance of semi-supervised methods is heavily dependent on the availability of unannotated data, and their effectiveness declines when such data are scarce or absent. To overcome this limitation, we propose a simple, yet effective and computationally efficient approach for medical image segmentation that leverages only existing annotations. We propose BoundarySeg , a multi-task framework that incorporates organ boundary prediction as an auxiliary task to full organ segmentation, leveraging consistency between the two task predictions to provide additional supervision. This strategy improves segmentation accuracy, especially in low data regimes, allowing our method to achieve performance comparable to or exceeding state-of-the-art semi supervised approaches all without relying on unannotated data or increasing computational demands. Code will be released upon acceptance.

Deep learning models can generate virtual immunohistochemistry (IHC) stains from hematoxylin and eosin (H&E) images, offering a scalable and low-cost alternative to laboratory IHC. However, reliable evaluation of image quality remains a challenge as current texture- and distribution-based metrics quantify image fidelity rather than the accuracy of IHC staining. Here, we introduce an automated and accuracy grounded framework to determine image quality across sixteen paired or unpaired image translation models. Using color deconvolution, we generate masks of pixels stained brown (i.e., IHC-positive) as predicted by each virtual IHC model. We use the segmented masks of real and virtual IHC to compute stain accuracy metrics (Dice, IoU, Hausdorff distance) that directly quantify correct pixel - level labeling without needing expert manual annotations. Our results demonstrate that conventional image fidelity metrics, including Frechet Inception Distance (FID), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM), correlate poorly with stain accuracy and pathologist assessment. Paired models such as PyramidPix2Pix and AdaptiveNCE achieve the highest stain accuracy, whereas unpaired diffusion- and GAN-based models are less reliable in providing accurate IHC positive pixel labels. Moreover, whole-slide images (WSI) reveal performance declines that are invisible in patch-based evaluations, emphasizing the need for WSI-level benchmarks. Together, this framework defines a reproducible approach for assessing the quality of virtual IHC models, a critical step to accelerate translation towards routine use by pathologists. 

Multiscale coupling between cell-scale biology and tissue-scale mechanics is a promising approach for modeling disease growth. In such models, tissue-level growth and remodeling (G&R) are driven by cell-level signaling pathways and systems biology models, where each model operates at different scales. Herein, we generate multiscale G&R models to capture the associated multiscale connections. At the cell-scale, we consider systems biology models in the form of systems of ordinary differential equations (ODEs) and partial differential equations (PDEs) representing the reactions between the biochemicals causing the growth based on mass-action or logic-based Hill-type kinetics. At the tissue-scale, we employ kinematic growth in continuum frameworks. Two illustrative test problems (a tissue graft and aneurysm growth) are examined with various chemical signaling networks, boundary conditions, and mechano-chemical coupling strategies. We extend two open-source software frameworks—febio and fenics—to disseminate examples of multiscale growth and remodeling simulations. One-way and two-way coupling between the systems biology and the growth models are compared and the effect of biochemical diffusivity and ODE versus PDE-based systems biology modeling on the G&R results are studied. The results show that growth patterns emerge from reactions between biochemicals, the choice between ODEs and PDEs systems biology modeling, and the coupling strategy. Cross-verification confirms that results for febio and fenics are nearly identical. We hope that these open-source tools will support reproducibility and education within the biomechanics community.

Recent developments in multimodal large language models (MLLM) have equipped language models to reason about vision and language jointly. This permits MLLMs to both perceive and answer questions about data visualization across a variety of designs and tasks. Applying MLLMs to a broad range of visualization tasks requires us to properly evaluate their capabilities, and the most common way to conduct evaluation is through measuring a model’s visualization reasoning capability, analogous to how we would evaluate human understanding of visualizations (e.g., visualization literacy). However, we found that in the context of visualization question answering (VisQA), how an MLLM perceives and reasons about visualizations can be fundamentally different from how humans approach the same problem. During the evaluation, even without visualization, the model could correctly answer a substantial portion of the visualization test questions, regardless of whether any selection options were provided. We hypothesize that the vast amount of knowledge encoded in the language model permits factual recall that supersedes the need to seek information from the visual signal. It raises concerns that the current VisQA evaluation may not fully capture the models’ visualization reasoning capabilities. To address this, we propose a comprehensive sanity check framework that integrates a rule-based decision tree and a sanity check table to disentangle the effects of ”seeing” (visual processing) and ”recall” (reliance on prior knowledge). This validates VisQA datasets for evaluation, highlighting where models are truly ”seeing”, positively or negatively affected by the factual recall, or relying on inductive biases for question answering. Our study underscores the need for careful consideration in designing future visualization understanding studies when utilizing MLLMs.

Idling vehicle detection (IVD) uses surveillance video and multichannel audio to localize and classify vehicles in the last frame as moving, idling, or engine-off in pick-up zones. IVD faces three challenges: (i) modality heterogeneity between visual cues and audio patterns; (ii) large box scale variation requiring multi-resolution detection; and (iii) training instability due to coupled detection heads. The previous end-to-end (E2E) model [1] with simple CBAM-based [2] bi-modal attention fails to handle these issues and often misses vehicles. We propose HAVT-IVD, a heterogeneity-aware network with a visual feature pyramid and decoupled heads. Experiments show HAVT-IVD improves mAP by 7.66 over the disjoint baseline and 9.42 over the E2E baseline.

Predicting particle trajectories with neural networks (NNs) has substantially enhanced many scientific and engineering domains. However, effectively quantifying and visualizing the inherent uncertainty in predictions remains challenging. Without an understanding of the uncertainty, the reliability of NN models in applications where trustworthiness is paramount is significantly compromised. This paper introduces the uncertainty tube, a novel, computationally efficient visualization method designed to represent this uncertainty in NN-derived particle paths. Our key innovation is the design and implementation of a superelliptical tube that accurately captures and intuitively conveys nonsymmetric uncertainty. By integrating well-established uncertainty quantification techniques, such as Deep Ensembles, Monte Carlo Dropout (MC Dropout), and Stochastic Weight Averaging-Gaussian (SWAG), we demonstrate the practical utility of the uncertainty tube, showcasing its application on both synthetic and simulation datasets.

Idling vehicle detection (IVD) uses surveillance video and multichannel audio to localize and classify vehicles in the last frame as moving, idling, or engine-off in pick-up zones. IVD faces three challenges: (i) modality heterogeneity between visual cues and audio patterns; (ii) large box scale variation requiring multi-resolution detection; and (iii) training instability due to coupled detection heads. The previous end-to-end (E2E) model with simple CBAM-based bi-modal attention fails to handle these issues and often misses vehicles. We propose HAVT-IVD, a heterogeneity-aware network with a visual feature pyramid and decoupled heads. Experiments show HAVT-IVD improves mAP by 7.66 over the disjoint baseline and 9.42 over the E2E baseline. 

While many of recent electoral reforms in the U.S. states have been centered on more restrictive legislation such as a higher standard of voter ID law, the call for expansive legal reforms has also drawn increasing attention. Among the expansive legislative proposals, arguably the most controversial one is to make voting compulsory. As Massachusetts learned from Australia and adopted the secret ballot in 1888, the Australian compulsory voting system which was enacted in 1924 has also been proposed recently by some U.S. legislators at the state level. The major supporters of compulsory voting in the U.S. have mainly come from the political left whose agendas have met vigorous opposition from the political right. Many criticisms are based on the assumptions of its beneficial effects on the voters at the margins whose voting participation has been historically lower than those of more well-to-do, informed voters, which would suggest a larger advantage for the Democratic Party. No empirical research, however, has shown such a possible Democratic advantage if compulsory voting were adopted. The main reason for no prior empirical support was because compulsory voting has never been implemented in the U.S. at the national level, and no previous studies have had comprehensive tools at hand to handle many layers and complexities in predicting the outcomes of compulsory voting at both the national and state levels. This article takes advantage of the most recent machine learning tools to narrow this empirical gap significantly. The ANES data spanning presidential elections from 1948 to 2020 provide the most comprehensive and unprecedented big data sets to test whether it is indeed the Democratic Party that would reap more benefits in compulsory voting at both the national and state levels. Our starting point is to use a state-of-the-art adaptation of the classic logit regression – elastic-net regression – for both our data-driven feature selection and modeling tasks. Furthermore, based on an innovative Transfer Component Analysis (TCA) approach added to our modeling pipeline, our machine learning prediction model shows that the effect of compulsory voting does benefit overall the Democratic Party at the national level. At the state level, however, the Republican Party can also achieve many positive electoral impacts within the context of compulsory voting.

Digital twins are developed to model the behavior of a specific physical asset (or twin), and they can consist of high-fidelity physics-based models or surrogates. A highly accurate surrogate is often preferred over multi-physics models as they enable forecasting the physical twin future state in real-time. To adapt to a specific physical twin, the digital twin model must be updated using in-service data from that physical twin. Here, we extend Gaussian process (GP) models to include derivative data, for improved accuracy, with dynamic updating to ingest physical twin data during service. Including derivative data, however, comes at a prohibitive cost of increased covariance matrix dimension. We circumvent this issue by using a sparse GP approximation, for which we develop extensions to incorporate derivatives. Numerical experiments demonstrate that the prediction accuracy of the derivative-enhanced sparse GP method produces improved models upon dynamic data additions. Lastly, we apply the developed algorithm within a DT framework to model fatigue crack growth in an aerospace vehicle. 

Air quality impacts on human health are an increasing concern globally. Vehicle pollution is a particular concern because of its multiple adverse health effects, and discretionary vehicle idling contributes significantly to local-scale poor air quality. This study introduces a novel approach to traditional static (unchanging) anti-idling signage. Here, we demonstrate a system, called SmartAir, that provides dynamic social-norm messages to drivers coupled with information about idling status or vehicle emissions in the area. A machine learning algorithm with audio and video inputs determines vehicle idling status. Vehicle emissions are measured using a suite of low-cost air quality nodes. In this study, we show that the SmartAir system reduces idling time by 28.0% and local CO2 concentrations by 29.5% compared with background.

Numerical methods for radiative transfer play a key role in modern-day astrophysics and cosmology, including study of the inhomogeneous reionization process. In this context, ray tracing methods are well-regarded for accuracy but notorious for high computational cost. In this work, we extend the capabilities of the Nyx N-body / hydrodynamics code, coupling radiation to gravitational and gas dynamics. We formulate adaptive ray tracing as a novel series of filters and transformations that can be used with AMReX particle abstractions, simplifying implementation and enabling portability across Exascale GPU architectures. To address computational cost, we present a new algorithm for merging sources, which significantly accelerates computation once reionization is well underway. Furthermore, we develop a novel prescription for geometric overlap correction with low-density neighbor cells. We perform verification and validation against standard analytic and numerical test problems. Finally, we demonstrate scaling to up to 1024 nodes and 4096 GPUs running multiphysics cosmological simulations, with 4096^3 Eulerian gas cells, 4096^3 dark matter particles, and ray tracing on a 1024^3 coarse grid. For these full cosmological simulations, we demonstrate convergence in terms of reionization history and post-ionization Lyman-alpha forest flux. 

Data visualizations are typically not accessible to blind and low-vision (BLV) users. Automatically generating text descriptions offers an enticing mechanism for democratizing access to the information held in complex scientific charts, yet appropriate procedures for generating those texts remain elusive. Pursuing this issue, we study a single complex chart form: UpSet plots. UpSet Plots are a common way to analyze set data, an area largely unexplored by prior accessibility literature. By analyzing the patterns present in real-world examples, we develop a system for automatically captioning any UpSet plot. We evaluated the utility of our captions via semi-structured interviews with (N=11) BLV users and found that BLV users find them informative. In extensions, we find that sighted users can use our texts similarly to UpSet plots and that they are better than naive LLM usage.

Given the constantly increasing number of newly published vulnerabilities, manually assessing their scores (e.g., under the Common Vulnerability Scoring System) has become unfeasible. Recently, learning-based systems have been proposed to automatically predict vulnerability scores. Such systems use vulnerability indexing databases to train deep learning algorithms. However, their practical applicability has important limitations, including a high dependency on the quality and diversity of training data, and high computational requirements. In addition, vulnerability descriptions often do not follow the standard templates and are not rich enough with respect to the expected features. In this paper, we propose a novel architecture that takes advantage of both generative artificial intelligence and lightweight deep learning techniques to provide an efficient and effective solution for automated vulnerability scoring. Data extracted from the National Vulnerability Dataset is fed into a large language model layer, whose output (i.e., an augmented dataset) is then used in a lightweight fine-tuned BERTsmall layer. We provide the results of an extensive experimental assessment of the effect of both each layer of the architecture and end-to-end performances. The results suggest that the combination of GPT3.5-Turbo and BERTsmall provides the most effective accuracy-time trade-off. We also compare the performance of the proposed architecture with other LLMs, BERT models, and cutting-edge approaches. The results show good improvements in prediction quality also when compared to a recent technique that incorporates data from 66 different sources, including the NVD.