Purpose.Diffusion-weighted imaging (DWI) has significant value in disease diagnosis and treatment response monitoring, but its inherent low signal-to-noise ratio (SNR) severely affects image quality and quantification accuracy. Existing denoising techniques often blur important tissue boundary information when suppressing noise.Methods.This study proposes a band-limited implicit neural representation (BL-INR) framework for DWI denoising. The method introduces BL positional encoding based on the frequency response characteristics of the sinc function to restrict INR models from learning high-frequency noise while maintaining strong signal representation capabilities. Furthermore, multi-b-value DWI and structural MRI from the same patient are integrated as anatomical priors, exploiting the correlation between true signals and the statistical independence of noise to achieve effective denoising.Main Results.In clinical DWI data evaluation across four anatomical regions (brain, head and neck, abdomen, and pelvis), BL-INR's visualization results were superior to existing methods. Under extremely low SNR conditions (SNR = 1) in simulated noise experiments, BL-INR achieved a peak SNR of 35.44 and structural similarity index of 0.933, significantly outperforming other methods. Phantom denoising results showed that BL-INR achieved an average apparent diffusion coefficient value error of only4.57×10-5 mm2 s-1, the smallest among all methods.Significance.BL-INR provides a novel approach for DWI denoising by limiting the frequency of INR input positional encoding. Its self-supervised learning characteristics require no paired training data and allow convenient clinical application. The method enables the derivation of accurate diffusion parameters, providing a reliable foundation for DWI-based quantitative analysis with significant clinical application value.

Abstract Martian exploration imagery is severely degraded by spatially heterogeneous dust, diminishing contrast and obscuring critical terrain details essential for geological analysis and navigation. Martian image dust removal is critically important, yet remains challenged by the failure of Earth-derived physical models, absolute absence of real paired data, and unphysical artifacts from unconstrained unsupervised methods. In this paper, we propose a novel physics-aware dehazing framework integrating: 1) A Multistage Fog Flow Network (MSFF-Net), a haze flow field-guided multi-stage network featuring a HazeFlowAligner module for deformable cross-scale feature alignment to explicitly resolve spatial heterogeneity, and Deformable Enhanced Convolution (Deformable ECV) modules with dynamic receptive fields to balance global haze removal and local detail preservation; 2) a terrain-adaptive data synthesis method using a physics-constrained CycleGAN to generate high-fidelity paired data for network training, which autonomously learns terrain-dependent dust density distributions and incorporates differentiable physical degradation modeling-including Perlin noise for realistic texturing, adaptive blur, and intensity-variant noise attenuation. Extensive experiments demonstrate state-of-the-art performance in perceptual quality and physical consistency compared to existing methods, achieving up to 46.7% and 77.8% improvements in PSNR and SSIM, respectively, on Martian dusty image, significantly outperforming prior arts, and showing exceptional generalization on real Tianwen-1 imagery and cross-domain SOTS benchmarks. This work establishes a new paradigm for physically-grounded visual restoration in planetary exploration and significantly enhances the accuracy of downstream tasks such as mineral mapping and autonomous navigation. Keywords： Mars dust, haze flow，CycleGAN，remote sensing image, dust removal

Abstract Recently, Graph Convolutional Networks (GCNs)-based methods have demonstrated superior performance in fault diagnosis for their powerful capabilities in data relationship mining. However, existing GCNs rely on labeled data for training, yet obtaining such labels in real-world industrial settings is typically challenging and unrealistic. Besides, over-smoothing commonly occurs when GCNs layers increase and temporal dependencies have not been fully considered during feature learning. To address these issues, this paper proposes an end-to-end Multiscale Graph Convolution Clustering Network (MSGCCN) for fault diagnosis using unlabeled data. Firstly, the unlabeled data are constructed as multi-scale temporal graphs to establish the temporal dependency relationships among data samples at different scales. Secondly, the multiscale graph convolution gate recurrent unit (MSGCG)-based encoder is introduced to learn temporal dependency relationships across different receptive fields. The stacked MSGCG is capable of extracting multi-scale temporal features from the constructed graphs by incorporating multi-scale graph convolution operation into GRU. Meanwhile, a Temporal and Scale Feature Fusion (TSFF) module is developed to adaptively perform effective feature fusion along temporal and scale dimensions, enabling the extraction of comprehensive deep fault representations. Finally, an end-to-end deep clustering loss is employed to guide the optimization of model parameters, enabling precise clustering of fault patterns. To validate the effectiveness of the proposed method, extensive experiments were performed on two experimental bench datasets. The proposed method demonstrates superior performance in unsupervised fault diagnosis, achieving accuracy rates of 93.56%, 91.83%, and 97.23% on three subdatasets, respectively, outperforming all comparative methods.

Objective.During radiotherapy, the radiation dose delivered to circulating blood can result in radiation-induced lymphopenia, which is correlated with adverse clinical outcomes like lower survival. Increasingly complex models to simulate radiation dose delivery to circulating blood have been developed in response, and their inclusion during radiotherapy treatment planning has been suggested. However, performing full dynamic blood dose simulations which take into account temporal considerations such as blood flow dynamics and treatment delivery time during the iterative treatment planning process is currently infeasible. This work presents a quasi-instantaneous deep learning-based approach to estimate blood dose simulation results to allow for their inclusion during treatment planning.Approach.We used treatment planning computed tomography images and dose-volume histograms of 157 head-and-neck cancer patients to perform dynamic blood dose simulations (HEDOS). Subsequently, a deep neural network composed of fully-connected layers and a Transformer encoder was trained to estimate the blood dose distribution obtained from HEDOS, using the same inputs as HEDOS. We used 126 patients' data for training and internal validation and the remaining 31 patients' data for testing. To evaluate the proposed method, we calculated the Kullback-Leibler (KL) divergence between the prediction results and the ground truth data. Additionally, we compared the minimum dose delivered to 90% of the blood particles receiving the highest dose (D90%) to estimate the model's clinical efficacy.Main results.The average and standard deviation of KL divergence between the prediction and the ground truth were 0.099 and 0.092, respectively. The D90%calculated from the predicted distribution showed a mean-absolute-percentage error of 4.60% compared to the ground truth.Significance.A deep learning-based model capable of accurately and quasi-instantaneously predicting the results of dynamic blood dose simulations was developed, paving the way for the inclusion of dynamic blood dose simulations during radiotherapy treatment planning.

This study aims to provide a comprehensive model to assist investors in strategic decision-making amid market uncertainty. Global economic uncertainty characterized by cycles of stagflation and recession has recurred in history and is expected to recur until 2025. This condition encourages the importance of investment strategies that can protect asset values from economic pressures. This study uses a quantitative approach with forecasting methods and risk analysis based on time series data. The data used are daily gold and silver prices from the London Bullion Market Association (LBMA) in USD, collected over a two-year period, namely from January 3, 2023 to January 4, 2025. The data is secondary and obtained from the official LBMA website. The research stages begin with a literature study to understand relevant concepts and methods, followed by data collection, and continued with data preprocessing. The preprocessing stages include checking for outliers, handling missing values using the series mean method, and merging data for temporal consistency. For the forecasting process, the Fuzzy Time Series–Markov Chain method is used, which consists of several steps: the formation of universe and interval sets using the Sturges formula, the definition of fuzzy sets, the fuzzification process, the formation of Fuzzy Logical Relationships (FLR) and Fuzzy Logical Relationship Groups (FLRG), and the preparation of transition probability matrices. The forecasting results are obtained through the defuzzification process, which are then evaluated using the Mean Absolute Percentage Error (MAPE) indicator to assess the accuracy of the model. Risk analysis is carried out using the Value at Risk (VaR) approach using the Extreme Value Theory (EVT) method and the Generalized Pareto Distribution (GPD). The entire analysis process is carried out using Microsoft Excel and RStudio software to ensure accuracy and efficiency in data processing and statistical modeling. This study has succeeded in developing a hybrid Fuzzy Time Series–Markov Chain model to forecast precious metal prices, especially gold and silver, with a very high level of accuracy. Based on an evaluation of various training and testing data proportions, the best model was obtained at a 95:5 ratio, with MAPE values of 0.66% for gold and 1.18% for silver in the training data, and 0.55% and 0.94% in the testing data. These results indicate that the model is able to effectively capture historical price patterns and provide predictions close to the actual value.

Modern diffusion-based image generative models have made significant progress and become promising to enrich training data for the object detection task. However, the generation quality and the controllability for complex scenes containing multi-class objects and dense objects with occlusions remain limited. This paper presents ODGEN, a novel method to generate high-quality images conditioned on bounding boxes, thereby facilitating data synthesis for object detection. Given a domain-specific object detection dataset, we first fine-tune a pre-trained diffusion model on both cropped foreground objects and entire images to fit target distributions. Then we propose to control the diffusion model using synthesized visual prompts with spatial constraints and object-wise textual descriptions. ODGEN exhibits robustness in handling complex scenes and specific domains. Further, we design a dataset synthesis pipeline to evaluate ODGEN on 7 domain-specific benchmarks to demonstrate its effectiveness. Adding training data generated by ODGEN improves up to 25.3% mAP@.50:.95 with object detectors like YOLOv5 and YOLOv7, outperforming prior controllable generative methods. We also design an evaluation protocol based on COCO-2014 to validate the synthetic data of ODGEN in general domains and observe an advantage up to 5.6% in mAP@.50:.95 against existing methods. In addition, we employ a series of large-scale object detection datasets to train a general model named Stable Box Diffusion, which covers thousands of object categories in most common scenes.

Lewy Body Disease (LBD) is a common but understudied dementia that poses a significant public health burden. It shares similar clinical signs with Alzheimer's disease (AD), with both conditions progressing through stages of normal cognition, mild cognitive impairment, and dementia. A major obstacle in LBD diagnosis is data scarcity, which limits the effectiveness of deep learning models. In contrast, AD datasets are more abundant, offering a potential avenue for knowledge transfer. However, LBD and AD data are typically collected from different sites using varied machines and protocols, resulting in a distinct domain shift. To effectively leverage AD data while mitigating domain shift, we propose a Transferability Aware Transformer (TAT) that adapts knowledge from AD to enhance LBD diagnosis. Our method utilizes structural connectivity (SC) derived from structural MRI as training data. Built on the attention mechanism, TAT assigns high weights to disease-transferable features while suppressing domain-specific ones, effectively reducing domain shift and improving diagnostic accuracy on limited LBD data. The experimental results demonstrate the effectiveness of TAT. Our work serves as the first to explore domain adaptation from AD to LBD study under data scarcity and domain shift scenarios, providing a promising framework for domain-adaptive diagnosis of rare diseases.

Understanding and predicting molecular properties remains a central challenge in scientific machine learning, especially when training data are limited or task-specific supervision is scarce. We introduce the Molecular Equivariant Transformer...

MSG: Stealing data from pruned neural networks via malicious sparsity guidance.

Hypothesis spaces for deep learning.

Category-specific unlabeled data risk minimization for ultrasound semi-supervised segmentation.

Recent progress in vision Transformers exhibits great success in various tasks driven by the new spatial modeling mechanism based on dot-product self-attention. In this paper, we show that the key ingredients behind the vision Transformers, namely input-adaptive, long-range and high-order spatial interactions, can also be efficiently implemented with a convolution-based framework. We present the Recursive Gated Convolution (${\mathit{g}}^{\mathit{n}}$gnConv) that performs high-order spatial interactions with gated convolutions and recursive designs. The new operation is highly flexible and customizable, which is compatible with various variants of convolution and extends the two-order interactions in self-attention to arbitrary orders without introducing significant extra computation. ${\mathit{g}}^{\mathit{n}}$gn Conv can serve as a plug-and-play module to improve various vision Transformers and convolution-based models. Based on the proposed operation, we construct a new family of generic vision backbones for various visual modalities and tasks, including HorNet and HorFPN for image recognition, Hor3D for point cloud analysis, and HorCLIP for vision-language modeling. For image recognition, we propose HorNet as a stronger visual encoder, where we conduct extensive experiments on ImageNet classification, COCO object detection, and ADE20K semantic segmentation. HorNet outperforms Swin Transformers and ConvNeXt by a significant margin with similar overall architecture and training configurations. HorNet also shows favorable scalability to more training data and larger model sizes. Apart from image encoders, we also show ${\mathit{g}}^{\mathit{n}}$gnConv can be applied to task-specific decoders and consistently improve dense prediction performance with less computation. For point cloud analysis, we design Hor3D, demonstrating the efficacy of high-order interactions for unstructured point cloud data through experiments on challenging 3D semantic segmentation tasks in S3DIS and ScanNet V2. In vision-language modeling, our proposed HorCLIP surpasses mainstream Vision Transformer and ConvNeXt architectures with shorter training schedules on ImageNet zero-shot classification and shows remarkably higher performance on vision-language dense representation tasks on COCO Panoptic datasets. Our results demonstrate that ${\mathit{g}}^{\mathit{n}}$gnConv with high-order spatial interactions can be a new basic operation for visual modeling that effectively combines the merits of both vision Transformers and CNNs.

The synthetic Turn in healthcare AI: Promise and Peril.

Multi-site brain MRI heterogeneity caused by differences in scanner field strengths, acquisition protocols, and software versions poses a significant challenge for consistent analysis. Image-level harmonization, leveraging advanced learning methods, has attracted increasing attention. However, existing methods often rely on paired data (e.g., human traveling phantoms) for training, which are not always available. Some methods perform MRI harmonization by transferring target-style features to source images but require explicitly learning disentangled image styles (e.g., contrast) via encoder-decoder networks, which increases computational complexity. This paper presents an unpaired MRI harmonization (UMH) framework based on a new image style-guided diffusion model. UMH operates in two stages: (1) a coarse harmonizer that aligns multi-site MRIs to a unified domain via a conditional latent diffusion model while preserving anatomical content; and (2) a fine harmonizer that adapts coarsely harmonized images to a specific target using style embeddings derived from a pre-trained Contrastive Language-Image Pre-training (CLIP) encoder, which captures semantic style differences between the original MRIs and their coarsely-aligned counterparts, eliminating the need for paired data. By leveraging rich semantic style representations of CLIP, UMH avoids learning image styles explicitly, thereby reducing computation costs. We evaluate UMH on 4,123 MRIs from three distinct multi-site datasets, with results suggesting its superiority over several state-of-the-art (SOTA) methods across image-level comparison, downstream classification, and brain tissue segmentation tasks.

Deep learning models for efficient geotechnical predictions: reducing training effort and data requirements with transfer learning

Abstract We model shortwave radiative transfer using a neural network that assigns physical meaning to the components within its structure. Separate components represent transmissivity, scattering, multilayer interreflection, and the spectral decomposition of radiation, where each of these contains computational paths for direct and diffuse radiation. This approach is an “open-box” alternative to a black-box neural network. It not only exposes the network’s internal variables to physical interpretation but also facilitates a synthesis of physical knowledge within the neural network. Beer’s law, the adding–doubling algorithm, conservation laws, the nonnegativity of physical variables, and physical independence are hardcoded into the network architecture. The resulting network contains just 3475 weights. We train it as a single system from input to output, jointly learning all weight values under the constraints of the embedded physical models and physical laws. Training and testing data were generated using the ecRad radiation scheme on input reanalysis data for eight atmospheric constituents from the Copernicus Atmosphere Monitoring Service. The network was trained on a dataset from 2008 and tested on datasets from 2009, 2015, and 2020, with error rates no worse than 0.026 K day −1 root-mean-square error (RMSE) in heating rate and 0.098 W m−2 RMSE in radiative flux, improving upon the prior state of the art in machine learning–based shortwave radiative transfer. Significance Statement Traditional physics-based modeling and data-driven machine learning often appear to be incompatible approaches for representing physical processes. We propose a synthesis that leverages the unique strengths of each. We overcome incomplete physics-based models by selectively incorporating data-driven neural network elements. Conversely, physical knowledge guides the design of the network’s architectureand its data-driven elements. The resulting network is not a typical black box but an “open box,” where its internal elements have well-defined physical functions and abide by physical laws. We demonstrate this novel approach by applying it to shortwave radiative transfer, a vital component of weather and climate forecasting models.

ABSTRACT Pipeline leakages have consistently been a major issue in district heating networks. Traditional hydraulic models face limitations in real-time dynamic mapping and sensor data reliability. Moreover, with technological advancement, integration with the Industrial Internet of Things (IIoT) has become a key consideration in the leakage detection process. In this study, we propose a leakage detection method that integrates a hydraulic model with IIoT to identify leaks in specific pipeline segments. Considering the measurement errors of pressure sensors, the pressure distribution data were corrected. In the optimal model, the trained pressure data achieved a prediction accuracy of 98.68% and an F1-macro score of 98.63%. The analysis of different sensor errors in this study contributes to improving leakage detection accuracy, while the integration with IIoT is also crucial for the future development of intelligent monitoring in heating networks.

Multi-annotation agreement and prediction consistency networks: Improving semi-supervised segmentation of medical images with ambiguous boundaries.

Assessment of a Zero-Shot Large Language Model in Measuring Documented Goals-of-Care Discussions.

We propose UAD, an end-to-end framework with Unsupervised pretext task for vision-based Autonomous Driving, achieving the best open-loop evaluation performance in nuScenes, meanwhile showing robust closed-loop driving quality in CARLA. Our motivation stems from the observation that current end-to-end autonomous driving (E2EAD) models still mimic the modular architecture in typical driving stacks, with carefully designed supervised perception and prediction subtasks to provide environment information for oriented planning. Although achieving groundbreaking progress, such design has certain drawbacks: 1) preceding subtasks require massive high-quality 3D annotations as supervision, posing a significant impediment to scaling the training data; and 2) each submodule entails substantial computation overhead in both training and inference. To this end, we propose UAD, an E2EAD framework with an unsupervised1 proxy to address all these issues. Firstly, we design a novel Angular Perception Pretext to eliminate the annotation requirement. The pretext perceives the driving scene by predicting the angular-wise spatial objectness and temporal dynamics, without manual annotation. Secondly, a self-supervised training strategy, which learns the consistency of the predicted trajectories under different augment views, is proposed to enhance the planning robustness in steering scenarios. Our UAD achieves 38.7% relative improvements over UniAD on the average collision rate of nuScenes open-loop evaluation and obtains the route completion score of 98.5% in closed-loop evaluation of CARLA's Town05 Long benchmark, which outperforms the recent work VADv2. Moreover, the proposed method consumes only 44.3% training resources of UniAD and runs $3.4\times$3.4× faster in inference when employing the same backbone network. Our innovative design not only for the first time demonstrates unarguable performance advantages over supervised counterparts, but also enjoys unprecedented efficiency in data, training, and inference.