Simulated Datasets Research Articles

Fast motion-compensated reconstruction for 4D-CBCT using deep learning-based groupwise registration.

Previous work has that deep learning (DL)-enhanced 4D cone beam computed tomography (4D-CBCT) images improve motion modeling and subsequent motion-compensated (MoCo) reconstruction for 4D-CBCT. However, building the motion model at treatment time via conventional deformable image registration (DIR) methods is not temporally feasible. This work aims to improve the efficiency of 4D-CBCT MoCo reconstruction using DL-based registration for the rapid generation of a motion model prior to treatment. 
Approach. An artifact-reduction DL model was first used to improve the initial 4D-CBCT reconstruction by reducing streaking artifacts. Based on the artifact-reduced phase images, a groupwise DIR employing DL was used to estimate the inter-phase motion model. Two DL DIR models using different learning strategies were employed: 1) a patient-specific one-shot DIR model which was trained from scratch only using the images to be registered, and 2) a population DIR model which was pre-trained using collected 4D-CT images from 35 patients. The registration accuracy of two DL DIR models was assessed and compared to a conventional groupwise DIR approach implemented in the Elastix toolbox using the publicly available DIR-Lab dataset, a Monte Carlo simulation dataset from the SPARE challenge, and two clinical cases. 
Main results. The patient-specific DIR model and the population DIR model demonstrated registration accuracy comparable to the conventional state-of-the-art methods on the DIR-Lab dataset. No significant difference in image quality was observed between the final MoCo reconstructions using the patient-specific model and population model for motion modeling, compared to using the conventional approach. The average runtime (hh:mm:ss) of the entire MoCo reconstruction on SPARE dataset was reduced from 01:37:26 using conventional DIR method to 00:10:59 using patient-specific model and 00:01:05 using the pre-trained population model. 
Significance. DL-based registration methods can improve the efficiency in generating motion models for 4D-CBCT without compromising the performance of final MoCo reconstruction.

Inference with Non-Homogeneous Lognormal Diffusion Processes Conditioned on Nearest Neighbor

In this work, we approach the forecast problem for a general non-homogeneous diffusion process over time with a different perspective from the classical one. We study the main characteristic functions as mean, mode, and α-quantiles conditioned on a future time, not conditioned on the past (as is normally the case), and we observe the specific formula in some interesting particular cases, such as Gompertz, logistic, or Bertalanffy diffusion processes, among others. This study aims to enhance classical inference methods when we need to impute data based on available information, past or future. We develop a simulation and obtain a dataset that is closer to reality, where there is no regularity in the number or timing of observations, to extend the traditional inference method. For such data, we propose using characteristic functions conditioned on the past or the future, depending on the closest point at which we aim to perform the imputation. The proposed inference procedure greatly reduces imputation errors in the simulated dataset.

Modeling High-Pressure Hydrogen Gas Leakages With Graph Neural Networks

Abstract The introduction of hydrogen-methane blends as fuel in gas turbines rises concerns on the capability of state of art ventilation systems to dilute possible fuel leaks in the enclosures. Traditional numerical methods to perform leak analysis are limited by the number of factors involved, i.e. location and direction of the leak, cross section area, gas pressure in the pipelines, gas composition, and location of external objects. Hence, this arise the need for novel and fast tools capable for the accurate prediction of fuel dispersion in leak scenarios. To this extent, we propose a novel machine learning approach to model gas leaks. The model is trained on a dataset of numerical simulations accounting for several hydrogen/ methane concentrations in the fuel, different storage to ambient pressure ratios at the leak section, and a set of cross-flow ventilation velocities. The architecture of the machine learning model is based on graph neural networks, to solve a node-level regression task predicting fuel concentration in space for different high pressure leak scenarios. The model shows a significant speed-up in predicting fuel dispersion with respect to conventional methodology (0.1 s vs 3.5 h) but the GPU memory requirements proved to be a problem when dealing with 3D domains.

Single-Scene SAR Image Data Augmentation Based on SBR and GAN for Target Recognition

High-performance neural networks for synthetic aperture radar (SAR) automatic target recognition (ATR) often encounter the challenge of data scarcity. The lack of sufficient labeled SAR image datasets leads to the consideration of using simulated data to supplement the dataset. On the one hand, electromagnetic computation simulations provide high amplitude accuracy but are inefficient for large-scale datasets due to their complex computations and physical models. On the other hand, ray tracing simulations offer high geometric accuracy and computational efficiency but struggle with low amplitude correctness, hindering accurate numerical feature extraction. Furthermore, the emergence of generative adversarial networks (GANs) provides a way to generate simulated datasets, trying to balance computational efficiency with image quality. Nevertheless, the simulated SAR images generated based on random noise lack constraints, and it is also difficult to generate images that exceed the parameter conditions of the real image’s training set. Hence, it is essential to integrate physics-based simulation techniques into GANs to enhance the generalization ability of the imaging parameters. In this paper, we present the SingleScene-SAR Simulator, an efficient framework for SAR image simulation that operates under limited real SAR data. This simulator integrates rasterized shooting and bouncing rays (SBR) with cycle GAN, effectively achieving both amplitude correctness and geometric accuracy. The simulated images are appropriate for augmenting datasets in target recognition networks. Firstly, the SingleScene-SAR Simulator employs a rasterized SBR algorithm to generate radar cross section (RCS) images of target models. Secondly, a specific training pattern for cycle GAN is established to translate noisy RCS images into simulated SAR images that closely resemble real ones. Finally, these simulated images are utilized for data augmentation. Experimental results based on the constructed dataset show that with only one scene SAR image containing 30 target chips, the SingleScene-SAR Simulator can efficiently produce simulated SAR images that exhibit high similarity in both spatial and statistical distributions compared with real images. By employing simulated SAR images for data augmentation, the accuracy of target recognition networks can be consistently and significantly enhanced.

Analysis of a Single Cell RNA-seq Workflow by Random Matrix Theory Methods.

Single cell RNA-seq (scRNAseq) workflows typically start with a count matrix and end with the clustering of sampled cells. While a range of methods have been developed to cluster scRNAseq datasets, no theoretical tools exist to explain why a particular cluster exists or why a hypothesized cluster is missing. Recently, several authors have shown that eigenvalues of scRNAseq count matrices can be approximated using random matrix models. In this work, we extend these previous works to the study of a scRNAseq workflow. We model scaled count matrices using random matrices with normally distributed entries. Using these random matrix models, we quantify the differential expression of a cluster and develop predictions for the workflow, and in particular clustering, as a function of the differential expression. We also use results from random matrix theory (RMT) to develop predictive formulas for portions of the scRNAseq workflow. Using simulated and real datasets, we show that our predictions are accurate if certain conditions hold on differential expression, with our RMT based predictions requiring particularly stringent condition. We find that real datasets violate these conditions, leading to bias in our predictions, but our predictions are better than a naive estimator and we point out future work that can improve the predictions. To our knowledge, our formulas represents the first predictive results for scRNAseq workflows.

Characterizing the Nanohertz Gravitational Wave Background Using a t-process Power Spectral Density

Pulsar timing arrays (PTAs) are sensitive to low-frequency gravitational waves (GWs), which induce correlated changes in millisecond pulsars’ timing residuals. PTA collaborations around the world have recently announced evidence of a nanohertz gravitational wave background (GWB), which may be produced by a population of supermassive black hole binaries (SMBHBs). The GWB is often modeled as following a power-law power spectral density (PSD); however, a GWB produced by a cosmological population of SMBHBs is expected to have a more complex power spectrum due to the discrete nature of the sources. In this paper, we investigate using a t-process PSD to model the GWB, which allows us to fit for both the underlying power-law amplitude and spectral index as well as deviations from that power law, which may be produced by individual nearby binaries. We create simulated data sets based on the properties of the NANOGrav 15 yr data set, and we demonstrate that the t-process PSD can accurately recover the PSD when deviations from a power law are present. With longer timed data sets and more pulsars, we expect the sensitivity of our PTAs to improve, which will allow us to precisely measure the PSD of the GWB and study the sources producing it.

Deep Learning Assisted Plasmonic Dark-Field Microscopy for Super-Resolution Label-Free Imaging.

Dark-field microscopy (DFM) is a widely used imaging tool, due to its high-contrast capability in imaging label-free specimens. Traditional DFM requires optical alignment to block the oblique illumination, and the resolution is diffraction-limited to the wavelength scale. In this work, we present deep-learning assisted plasmonic dark-field microscopy (DAPD), which is a single-frame super-resolution method using plasmonic dark-field (PDF) microscopy and deep-learning assisted image reconstruction. Specifically, we fabricated a designed PDF substrate with surface plasmon polaritons (SPPs) illuminating specimens on the substrate. Dark field images formed by scattered light from the specimen are further processed by a pretrained convolutional neural network (CNN) using a simulation dataset based on the designed substrate and parameters of the detection optics. We demonstrated a resolution enhancement of 2.8 times on various label-free objects with a large potential for future improvement. We highlight our technique as a compact alternative to traditional DFM with a significantly enhanced spatial resolution.

Outlier Detection in Bayesian Neural Networks

Describing uncertainty is one of the major issues in modern deep learning. Artificial Intelligence models could be used with greater confidence by having solid methods for identifying and quantifying uncertainty. This article proposes two alternative methods for outlier detection in Bayesian Neural Networks used for classification tasks. This is done by looking for unusual pre-activation neuron values in the last layer of a Bayesian Neural Network. The proposed methods are compared to a baseline method for outlier detection, Predictive Entropy, on three datasets: a simulated dataset, the MNIST dataset, and the Breast Cancer Wisconsin dataset. In addition, we introduce a method for separating In-Between and Out-Of-Distribution outliers. The results indicate that the proposed methods' performance depends on the dataset type. In the case of a simple simulated dataset, we see that the proposed methods outperform the baseline method. The proposed method also outperforms the baseline method on the Breast Cancer Wisconsin dataset. However, when tested on the MNIST image dataset, we observed that the baseline is better than the proposed method. As expected, for all three datasets, an OR-combination of the two methods gives slightly better POWER score than the two methods separately, but this comes at the price of also obtaining a higher FDR score.

Comparative Evaluation of Open-Source Bioinformatics Pipelines for Full-Length Viral Genome Assembly.

The increasingly widespread application of next-generation sequencing (NGS) in clinical diagnostics and epidemiological research has generated a demand for robust, fast, automated, and user-friendly bioinformatics workflows. To guide the choice of tools for the assembly of full-length viral genomes from NGS datasets, we assessed the performance and applicability of four open-source bioinformatics pipelines (shiver-for which we created a user-friendly Dockerized version, referred to as dshiver; SmaltAlign; viral-ngs; and V-pipe) using both simulated and real-world HIV-1 paired-end short-read datasets and default settings. All four pipelines produced consensus genome assemblies with high quality metrics (genome fraction recovery, mismatch and indel rates, variant calling F1 scores) when the reference sequence used for assembly had high similarity to the analyzed sample. The shiver and SmaltAlign pipelines (but not viral-ngs and V-Pipe) also showed robust performance with more divergent samples (non-matching subtypes). With empirical datasets, SmaltAlign and viral-ngs exhibited an order of magnitude shorter runtime compared to V-Pipe and shiver. In terms of applicability, V-Pipe provides the broadest functionalities, SmaltAlign and dshiver combine user-friendliness with robustness, while the use of viral-ngs requires less computational resources compared to other pipelines. In conclusion, if a closely matched reference sequence is available, all pipelines can reliably reconstruct viral consensus genomes; therefore, differences in user-friendliness and runtime may guide the choice of the pipeline in a particular setting. If a matched reference sequence cannot be selected, we recommend shiver or SmaltAlign for robust performance. The new Dockerized version of shiver offers ease of use in addition to the accuracy and robustness of the original pipeline.

DHDP-SLAM: Dynamic Hierarchical Dirichlet Process based data association for semantic SLAM

The key to a semantic SLAM system lies in the data association between measurements and landmarks, using the association results to provide constraints for the pose estimation of robot. However, to address the issues in data association models where data continuity and similarity are not sufficiently emphasized and single-level association strategies exhibit low robustness, we propose a data association method based on the Dynamic Hierarchical Dirichlet Process (DHDP), which is an online data association model that can make full use of the continuity and similarity between data to improve the convergence speed of the model, and at the same time, it can also dynamically take into account the influence of previous data on the current data. Additionally, DHDP has a more robust two-level association strategy to improve the accuracy of data association. In the experiments, three different datasets (Simulation dataset, KITTI dataset and TUM dataset) were selected to validate the proposed method, and the results show that DHDP has faster convergence speed and higher association accuracy, and it is able to provide additional constraints to the system when integrating it into the SLAM system, and by compared it with the state-of-the-art SLAM methods, the DHDP-SLAM exhibits higher localization accuracy.

Dual graph-regularized low-rank representation for hyperspectral image denoising

Hyperspectral images have a wide range of applications in many fields. However, when hyperspectral images are captured by spectrometers, there is inevitably considerable noise, which affects subsequent research. In recent years, many hyperspectral image denoising methods based on low-rank representations have been proposed. Artificial intelligence denoising methods are also popular. However, the research on multi noise denoising is rarely mentioned, and most literatures only focus on one noise in hyperspectral images. Thus, we propose a denoising model for hyperspectral image based on dual graph-regularized low-rank representation, which can not only reduce multiple types of noise simultaneously, but also preserves details of the original image. In particular, this is the first time that the dual low-rank representation and dual graph regularizations are used on hyperspectral images. We solve this method using the linearized alternating direction method with adaptive penalty. Finally, we conduct experiments on simulated and real data sets to verify the effectiveness of our method. The experimental results show that our method can not only effectively remove a variety of mixed noises, but also well retain the details of the image.

DeepHapNet: a haplotype assembly method based on RetNet and deep spectral clustering.

Gene polymorphism originates from single-nucleotide polymorphisms (SNPs), and the analysis and study of SNPs are of great significance in the field of biogenetics. The haplotype, which consists of the sequence of SNP loci, carries more genetic information than a single SNP. Haplotype assembly plays a significant role in understanding gene function, diagnosing complex diseases, and pinpointing species genes. We propose a novel method, DeepHapNet, for haplotype assembly through the clustering of reads and learning correlations between read pairs. We employ a sequence model called Retentive Network (RetNet), which utilizes a multiscale retention mechanism to extract read features and learn the global relationships among them. Based on the feature representation of reads learned from the RetNet model, the clustering process of reads is implemented using the SpectralNet model, and, finally, haplotypes are constructed based on the read clusters. Experiments with simulated and real datasets show that the method performs well in the haplotype assembly problem of diploid and polyploid based on either long or short reads. The code implementation of DeepHapNet and the processing scripts for experimental data are publicly available at https://github.com/wjj6666/DeepHapNet.

AI-Powered Biodiversity Assessment: Species Classification via DNA Barcoding and Deep Learning

Only 1.2 million out of an estimated 8.7 million species on Earth have been fully classified through taxonomy. As biodiversity loss accelerates, ecologists are urgently revising conservation strategies, but the “taxonomic impediment” remains a significant barrier, limiting effective access to and understanding of taxonomic data for many researchers. As sequencing technologies advance, short DNA sequence fragments increasingly serve as DNA barcodes for species identification. Rapid acquisition of DNA sequences from diverse organisms is now possible, highlighting the increasing significance of DNA sequence analysis tools in species identification. This study introduces a new approach for species classification with DNA barcodes based on an ensemble of deep neural networks (DNNs). Several techniques are proposed and empirically evaluated for converting raw DNA sequence data into images fed into the DNNs. The best-performing approach is obtained by representing each pair of DNA bases with the value of a related physicochemical property. By utilizing different physicochemical properties, we can create an ensemble of networks. Our proposed ensemble obtains state-of-the-art performance on both simulated and real datasets.

MCGAE: unraveling tumor invasion through integrated multimodal spatial transcriptomics.

Spatially Resolved Transcriptomics (SRT) serves as a cornerstone in biomedical research, revealing the heterogeneity of tissue microenvironments. Integrating multimodal data including gene expression, spatial coordinates, and morphological information poses significant challenges for accurate spatial domain identification. Herein, we present the Multi-view Contrastive Graph Autoencoder (MCGAE), a cutting-edge deep computational framework specifically designed for the intricate analysis of spatial transcriptomics (ST) data. MCGAE advances the field by creating multi-view representations from gene expression and spatial adjacency matrices. Utilizing modular modeling, contrastive graph convolutional networks, and attention mechanisms, it generates modality-specific spatial representations and integrates them into a unified embedding. This integration process is further enriched by the inclusion of morphological image features, markedly enhancing the framework's capability to process multimodal data. Applied to both simulated and real SRT datasets, MCGAE demonstrates superior performance in spatial domain detection, data denoising, trajectory inference, and 3D feature extraction, outperforming existing methods. Specifically, in colorectal cancer liver metastases, MCGAE integrates histological and gene expression data to identify tumor invasion regions and characterize cellular molecular regulation. This breakthrough extends ST analysis and offers new tools for cancer and complex disease research.

ADM: adaptive graph diffusion for meta-dimension reduction.

Dimension reduction is essential for analyzing high-dimensional data, with various techniques developed to address diverse data characteristics. However, individual methods often struggle to capture all intricate patterns and complex structures simultaneously. To overcome this limitation, we introduce ADM (Adaptive graph Diffusion for Meta-dimension reduction), a novel meta-dimension reduction method grounded in graph diffusion theory. ADM integrates results from multiple dimension reduction techniques, leveraging their individual strengths while mitigating their specific weaknesses.ADM utilizes dynamic Markov processes to transform Euclidean space results into an information space, revealing intrinsic nonlinear manifold structures that are hard to capture by conventional methods. A critical advancement in ADM is its adaptive diffusion mechanism, which dynamically selects optimal diffusion time scales for each sample, enabling effective representation of multi-scale structures. This approach generates robust, high-quality low-dimensional representations that capture both local and global data structures while reducing noise and technique-specific distortions. We demonstrate ADM's efficacy on simulated and real-world datasets, including various omics data types. Results show that ADM provides clearer separation between biological groups and reveals more meaningful patterns compared to existing methods, advancing the analysis and visualization of complex biological data.

UPicker: a semi-supervised particle picking transformer method for cryo-EM micrographs.

Automatic single particle picking is a critical step in the data processing pipeline of cryo-electron microscopy structure reconstruction. In recent years, several deep learning-based algorithms have been developed, demonstrating their potential to solve this challenge. However, current methods highly depend on manually labeled training data, which is labor-intensive and prone to biases especially for high-noise and low-contrast micrographs, resulting in suboptimal precision and recall. To address these problems, we propose UPicker, a semi-supervised transformer-based particle-picking method with a two-stage training process: unsupervised pretraining and supervised fine-tuning. During the unsupervised pretraining, an Adaptive Laplacian of Gaussian region proposal generator is proposed to obtain pseudo-labels from unlabeled data for initial feature learning. For the supervised fine-tuning, UPicker only needs a small amount of labeled data to achieve high accuracy in particle picking. To further enhance model performance, UPicker employs a contrastive denoising training strategy to reduce redundant detections and accelerate convergence, along with a hybrid data augmentation strategy to deal with limited labeled data. Comprehensive experiments on both simulated and experimental datasets demonstrate that UPicker outperforms state-of-the-art particle-picking methods in terms of accuracy and robustness while requiring fewer labeled data than other transformer-based models. Furthermore, ablation studies demonstrate the effectiveness and necessity of each component of UPicker. The source code and data are available at https://github.com/JachyLikeCoding/UPicker.

Machine learning visualization tool for exploring parameterized hydrodynamics**This manuscript has been authored by Lawrence Livermore National Security, LLC under Contract No. DE-AC52-07NA2 7344 with the US. Department of Energy. The United States Government retains, and the publisher, by accepting the article for publication, acknowledges that the United States Government retains a non-exclusive, paid-up, irrevocable, world-wide license to publish or reproduce

Abstract We are interested in the computational study of shock hydrodynamics, i.e. problems involving compressible solids, liquids, and gases that undergo large deformation. These problems are dynamic and nonlinear and can exhibit complex instabilities. Due to advances in high performance computing it is possible to parameterize a hydrodynamic problem and perform a computational study yielding Ο(TB) of simulation state data. We present an interactive machine learning tool that can be used to compress, browse, and interpolate these large simulation datasets. This tool allows computational scientists and researchers to quickly visualize "what-if" situations, perform sensitivity analyses, and optimize complex hydrodynamic experiments.

Real-time fatigue assessment of Floating Offshore Wind Turbine Mooring employing sequence-to-sequence-based deep learning on indirect fatigue response

Mooring lines in Floating Offshore Wind Turbines (FOWT) are subjected to significant environmental loading, leading to fatigue degradation over prolonged usage which necessitates real-time monitoring to ensure operational safety. Traditional fatigue assessment measures stress fluctuations under operational loads with expensive underwater sensors, which are costly and complex to install, especially for monitoring mooring at higher depths. This eventually increases the levelized cost of energy and maintenance expenses. This study proposes a cost-efficient fatigue monitoring method for catenary mooring using a novel sequence-to-sequence deep-learning (DL) approach that leverages platform motions as indirect fatigue response to infer fatigue assessment. By correlating FOWT platform motions with their impact on the fatigue life of mooring lines, the proposed method offers an alternative to costly underwater sensor measurements. In the absence of real data, the DL model is trained using simulated datasets for an OC4 semi-submersible wind turbine model in the OpenFAST open-source simulation package, capturing a wide range of tension fluctuations under diverse metocean conditions. Accordingly, this synthetic data has been treated as “real data” throughout this study. The trained network effectively captures the nonlinear relationship between platform motion and mooring fatigue response in real-time, enabling efficient assessment of fatigue-induced damage on mooring lines.

Non parametric observation-driven hidden Markov model

Hidden Markov models (HMM) are used in different fields to study the dynamics of a process that cannot be directly observed. However, in some cases, the structure of dependencies of a HMM is too simple to describe the dynamics of the hidden process. In particular, in some applications in finance and in ecology, the transition probabilities of the hidden Markov chain can also depend on the current observation. In this work, we are interested in extending the classical HMM to this situation. We refer to the extended model as the observation-driven hidden Markov model (OD-HMM). We present a complete study of the general non parametric OD-HMM with discrete and finite state spaces. We study its identifiability and the consistency of the maximum likelihood estimators. We derive the associated forward-backward equations for the E-step of the EM algorithm. The quality of the procedure is tested on simulated datasets. We illustrate the use of the model on an application focused on the study of annual plant dynamics. This work establishes theoretical and practical foundations for this framework that could be further extended to the parametric context in order to simplify estimation and to hidden semi-Markov models for more realism.

DeFFace: Deep Face Recognition Unlocked by Illumination Attributes

General face recognition is currently one of the key technologies in the field of computer vision, and it has achieved tremendous success with the support of deep-learning technology. General face recognition models currently exhibit extremely high accuracy on some high-quality face datasets. However, their performance decreases in challenging environments, such as low-light scenes. To enhance the performance of face recognition models in low-light scenarios, we propose a face recognition approach based on feature decoupling and fusion (DeFFace). Our main idea is to extract facial-related features from images that are not influenced by illumination. First, we introduce a feature decoupling network (D-Net) to decouple the image into facial-related features and illumination-related features. By incorporating the illumination triplet loss optimized with unpaired identity IDs, we regulate illumination-related features to minimize the impact of lighting conditions on the face recognition system. However, the decoupled features are relatively coarse. Therefore, we introduce a feature fusion network (F-Net) to further extract the residual facial-related features from the illumination-related features and fuse them with the initial facial-related features. Finally, we introduce a lighting-facial correlation loss to reduce the correlation between the two decoupled features in the specific space. We demonstrate the effectiveness of our method on four real-world low-light datasets and three simulated low-light datasets. We retrain multiple general face recognition methods using our proposed low-light training sets to further validate the advanced performance of our method. Compared to general face recognition methods, our approach achieves an average improvement of more than 2.11 percentage points on low-light face datasets. In comparison with image enhancement-based solutions, our method shows an average improvement of around 16 percentage points on low-light datasets, and it also delivers an average improvement of approximately 5.67 percentage points when compared to illumination normalization-based methods.