Deep Learning Methods Research Articles

Spatiotemporal Interaction Based Dynamic Adversarial Adaptive Graph Neural Network for Air-Quality Prediction

Air quality issues have become a major environmental concern, with severe air pollution significantly reducing air quality and posing threats to human health. Accurate air quality prediction is crucial for preventing individuals from suffering the detrimental effects of severe air pollution. Recently, deep learning methods based on spatiotemporal graph neural networks (GNNs) have made considerable progress in modeling the temporal and spatial dependencies within air quality data by integrating GNNs with sequential models. Unfortunately, previous work often treats temporal and spatial dependencies as independent components, neglecting the intricate interactions between them. This oversight prevents the models from fully exploiting the complex spatiotemporal dependencies in the data, adversely affecting their predictive performance. To address these issues, we propose a general spatiotemporal interaction framework for air quality prediction. This framework models the bidirectional interactions between temporal and spatial dependencies in a data-driven manner. Furthermore, we designed a spatiotemporal feature extraction module and a dynamic adversarial adaptive graph learning module based on this framework. We introduce the Spatial-Temporal Interaction based Dynamic Adversarial Adaptive Graph Neural Network, capable of capturing the complex interactions between spatiotemporal dependencies and learning the dynamic spatial topology among sites by incorporating the competitive optimization concept of generative adversarial networks. Extensive experiments on two real-world datasets demonstrate the effectiveness of the proposed method, outperforming existing baseline models.

Predicting metabolite response to dietary intervention using deep learning

Due to highly personalized biological and lifestyle characteristics, different individuals may have different metabolite responses to specific foods and nutrients. In particular, the gut microbiota, a collection of trillions of microorganisms living in the gastrointestinal tract, is highly personalized and plays a key role in the metabolite responses to foods and nutrients. Accurately predicting metabolite responses to dietary interventions based on individuals’ gut microbial compositions holds great promise for precision nutrition. Existing prediction methods are typically limited to traditional machine learning models. Deep learning methods dedicated to such tasks are still lacking. Here we develop a method McMLP (Metabolite response predictor using coupled Multilayer Perceptrons) to fill in this gap. We provide clear evidence that McMLP outperforms existing methods on both synthetic data generated by the microbial consumer-resource model and real data obtained from six dietary intervention studies. Furthermore, we perform sensitivity analysis of McMLP to infer the tripartite food-microbe-metabolite interactions, which are then validated using the ground-truth (or literature evidence) for synthetic (or real) data, respectively. The presented tool has the potential to inform the design of microbiota-based personalized dietary strategies to achieve precision nutrition.

Benchmarking protein language models for protein crystallization

The problem of protein structure determination is usually solved by X-ray crystallography. Several in silico deep learning methods have been developed to overcome the high attrition rate, cost of experiments and extensive trial-and-error settings, for predicting the crystallization propensities of proteins based on their sequences. In this work, we benchmark the power of open protein language models (PLMs) through the TRILL platform, a be-spoke framework democratizing the usage of PLMs for the task of predicting crystallization propensities of proteins. By comparing LightGBM / XGBoost classifiers built on the average embedding representations of proteins learned by different PLMs, such as ESM2, Ankh, ProtT5-XL, ProstT5, xTrimoPGLM, SaProt with the performance of state-of-the-art sequence-based methods like DeepCrystal, ATTCrys and CLPred, we identify the most effective methods for predicting crystallization outcomes. The LightGBM classifiers utilizing embeddings from ESM2 model with 30 and 36 transformer layers and 150 and 3000 million parameters respectively have performance gains by 3- than all compared models for various evaluation metrics, including AUPR (Area Under Precision-Recall Curve), AUC (Area Under the Receiver Operating Characteristic Curve), and F1 on independent test sets. Furthermore, we fine-tune the ProtGPT2 model available via TRILL to generate crystallizable proteins. Starting with 3000 generated proteins and through a step of filtration processes including consensus of all open PLM-based classifiers, sequence identity through CD-HIT, secondary structure compatibility, aggregation screening, homology search and foldability evaluation, we identified a set of 5 novel proteins as potentially crystallizable.

Fast prediction of curing residual stress in polymer‐matrix composites by a deep learning method

AbstractCuring residual stress (CRS) is common in polymer‐matrix composites due to the anisotropic properties of materials. Finite element method (FEM), the most extensively used approach for curing behavior prediction, is usually complicated and time‐consuming. To achieve a fast prediction of process‐induced stresses, a convolutional neural network (CNN) is established based on the FEM. Firstly, a fully coupled methodology is built and validated through distortions of experimentally manufactured laminates. Then, it is applied to generate models with different stacking layers, and the residual stress is computed and serves as the dataset of the deep learning model. Finally, the construction and hyperparameters are determined, and the good generalization performance proves the high accuracy of the current model. Besides, the CNN method (<1 s) greatly reduces the computational time compared with the FEM (>14 min). The supervised machine learning method shows great potential in promoting efficiency in stacking sequence designing and optimization of composites.Highlights A fully coupled model was built to predict the curing behavior of composites. The numerical model was validated by the deformation of unsymmetrical composites. A FEM‐CNN method was proposed for the fast prediction of curing residual stress. Compared with FEM, this machine learning method is accurate and efficient.

Comparative study of CNN-based and conventional fault interpretation methods: a study of the deep-water Orange Basin, South Africa

Image-based deep learning methods, especially convolutional neural networks (CNNs), are gaining traction in seismic interpretation, but their application still demands manual validation. This study compares a U-Net structured CNN, called Fault-Net, with conventional edge-enhancing seismic attributes of variance and chaos, which serve as a scientific baseline. We adopted two seismic interpretation workflows: (1) conventional attributes to enhance fault features, and (2) a deep learning-based workflow for fault segmentation using CNNs. Both workflows were applied to a high-resolution 3D seismic dataset from the structurally complex deep-water Orange Basin (South Africa). While deep learning-based software packages are commercially available, it is unclear whether they are suitable for the Orange Basin and in an academic setting due to their proprietary architectures and generally closed training data. This study provides public evidence of the feasibility of automated structural interpretation in complex seismic datasets using deep learning, revealing both key benefits and limitations. Where high-quality labelled data are available, the deep learning approach is faster and tends to produce a cleaner and more accurate depiction of larger faults compared to conventional methods. The open availability of Fault-Net makes deep learning-based interpretation particularly advantageous for academic settings, offering significant time and resource efficiency while enhancing the understanding of complex subsurface structures.

Overview of Deep Learning and Nondestructive Detection Technology for Quality Assessment of Tomatoes

Tomato, as the vegetable queen, is cultivated worldwide due to its rich nutrient content and unique flavor. Nondestructive technology provides efficient and noninvasive solutions for the quality assessment of tomatoes. However, processing the substantial datasets to achieve a robust model and enhance detection performance for nondestructive technology is a great challenge until deep learning is developed. The aim of this paper is to provide a systematical overview of the principles and application for three categories of nondestructive detection techniques based on mechanical characterization, electromagnetic characterization, as well as electrochemical sensors. Tomato quality assessment is analyzed, and the characteristics of different nondestructive techniques are compared. Various data analysis methods based on deep learning are explored and the applications in tomato assessment using nondestructive techniques with deep learning are also summarized. Limitations and future expectations for the quality assessment of the tomato industry by nondestructive techniques along with deep learning are discussed. The ongoing advancements in optical equipment and deep learning methods lead to a promising outlook for the application in the tomato industry and agricultural engineering.

Mapping the Green Urban: A Comprehensive Review of Materials and Learning Methods for Green Infrastructure Mapping

Green infrastructure (GI) plays a crucial role in sustainable urban development, but effective mapping and analysis of such features requires a detailed understanding of the materials and state-of-the-art methods. This review presents the current landscape of green infrastructure mapping, focusing on the various sensors and image data, as well as the application of machine learning and deep learning techniques for classification or segmentation tasks. After finding articles with relevant keywords, the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyzes) method was used as a general workflow, but some parts were automated (e.g., screening) by using natural language processing and large language models. In total, this review analyzed 55 papers that included keywords related to GI mapping and provided materials and learning methods (i.e., machine or deep learning) essential for effective green infrastructure mapping. A shift towards deep learning methods can be observed in the mapping of GIs as 33 articles use various deep learning methods, while 22 articles use machine learning methods. In addition, this article presents a novel methodology for automated verification methods, demonstrating their potential effectiveness and highlighting areas for improvement.

Traffic incident duration prediction based on deep learning methods

ABSTRACT Traffic congestion increases travel time and pollution, necessitating precise incident handling time predictions. This study addresses data imbalance, feature granularity, and label inconsistencies using K-means clustering, re-labelling, and concept hierarchy refinements. A predictive framework integrates classification and regression models to estimate incident duration for various accident types. Results show Random Forest achieves 94.32% accuracy for clearance time, XGBoost a mean absolute error (MAE) of 6.44 minutes for response time, and Support Vector Machine with Gradient Boosting Regression Trees a MAE of 13.3 minutes for total incident duration. These results support quick countermeasures, reduce congestion, and inform drivers via highway signs.

Remote sensing image destriping with two-stage image decomposition network

ABSTRACT Stripe noise often affects remote sensing images, leading to the degradation of imaging quality and impacting subsequent image processing. Deep learning methods have made remarkable advancements in removing stripes from remote sensing images with their powerful feature extraction capability. However, it is noteworthy that these methods are still insufficient in analysing structural characteristics in the direction of stripes and multi-scale contextual information. To deal with the above issues, we propose a remote sensing image destriping with a two-stage image decomposition network, named TSIDNet, which utilizes the stripe structural characteristics to effectively remove stripe noise while retaining better details. Firstly, stripes are directional, and the structural information of the stripes is mainly concentrated in the image after the differential decomposition of the direction along the stripes. Therefore, a differential decomposition stripe extraction subnetwork (DDSESN) is constructed to generate latent images. Within this subnetwork, we further design a multi-scale cross-fusion residual block (MCRB) and cross-scale fusion attention block (CSFAB) to gradually expand the network’s receptive field, which is conducive to extract and remove stripes more completely. Secondly, the features obtained by adding the latent clear image with details extracted from the stripe’s reverse direction are input into a DWT decomposition detail enhancement subnetwork (DDDESN), which utilizes the discrete wavelet transform (DWT) and the dense residual network for detail enhancement. Demonstrated through experiments, the proposed TSIDNet enables the removal of image stripes while preserving details, and surpassing the comparative methods in qualitative as well as quantitative evaluations. The code will be provided after acceptance.

Three-dimensional semi-supervised lumbar vertebrae region of interest segmentation based on MAE pre-training

Background: The annotation of the regions of interest (ROI) of lumbar vertebrae by radiologists for bone density assessment is a tedious and time-intensive task. However, deep learning (DL) methods for image segmentation has the potential to substitute manual annotations which can significantly improve the efficiency of clinical diagnostics. Objective: The paper proposes a semi-supervised three-dimensional (3D) segmentation method for the ROI of lumbar vertebrae by integrating the tube masking masked autoencoder (MAE) pre-training. Methods: The paper proposes a method that modifies the masking strategy of the original MAE pre-training network. And the pre-training network is only trained by images without segmentation labels, when the training is finished, the weights will be saved for segmentation tasks. In downstream tasks, a semi-supervised approach utilizing pseudo-label generation is employed for training. This method leverages a small amount of labeled data to achieve the segmentation of ROI of the lumbar vertebrae. Results: The experimental results demonstrate that under the condition of limited annotated data, the proposed network improves the dice coefficient by 5–7% and reduces the hausdorff distance by 0.2∼0.6 mm compared to using the UNetr network alone for segmentation. When compared to the conventional MAE, the tube masking MAE presented in this paper assists effectively in segmentation, resulting in a 2% increase in the dice coefficient and a 0.24 mm reduction in the hausdorff distance. Conclusion: Automatic segmentation of the ROI of the lumbar vertebrae helps to shorten the time for doctors to annotate vertebrae during clinical bone density examinations. The paper employs the tube masking MAE pre-trained model to effectively extract contextual information of the 3D lumbar vertebrae, combining it with a semi-supervised network leveraging pseudo-label generation for fine-tuning, which leads to effective 3D segmentation of the lumbar vertebrae.

Heterogeneous entity representation for medicinal synergy prediction.

Forecasting the synergistic effects of drug combinations facilitates drug discovery and development, especially regarding cancer therapeutics. While numerous computational methods have emerged, most of them fall short in fully modeling the relationships among clinical entities including drugs, cell lines, and diseases, which hampers their ability to generalize to drug combinations involving unseen drugs. These relationships are complex and multidimensional, requiring sophisticated modeling to capture nuanced interplay that can significantly influence therapeutic efficacy. We present a novel deep hypergraph learning method named Heterogeneous Entity Representation for MEdicinal Synergy prediction (HERMES) to predict the synergistic effects of anti-cancer drugs. Heterogeneous data sources, including drug chemical structures, gene expression profiles, and disease clinical semantics, are integrated into hypergraph neural networks equipped with a gated residual mechanism to enhance high-order relationship modeling. HERMES demonstrates state-of-the-art performance on two benchmark datasets, significantly outperforming existing methods in predicting the synergistic effects of drug combinations, particularly in cases involving unseen drugs. The source code is available at https://github.com/Christina327/HERMES. Supplementary data are available at Bioinformatics online.

SVDDD: SAR Vehicle Target Detection Dataset Augmentation Based on Diffusion Model

In the field of target detection using synthetic aperture radar (SAR) images, deep learning-based supervised learning methods have demonstrated outstanding performance. However, the effectiveness of deep learning methods is largely influenced by the quantity and diversity of samples in the dataset. Unfortunately, due to various constraints, the availability of labeled image data for training SAR vehicle detection networks is quite limited. This scarcity of data has become one of the main obstacles hindering the further development of SAR vehicle detection. In response to this issue, this paper collects SAR images of the Ka, Ku, and X bands to construct a labeled dataset for training Stable Diffusion and then propose a framework for data augmentation for SAR vehicle detection based on the Diffusion model, which consists of a fine-tuned Stable Diffusion model, a ControlNet, and a series of methods for processing and filtering images based on image clarity, histogram, and an influence function to enhance the diversity of the original dataset, thereby improving the performance of deep learning detection models. In the experiment, the samples we generated and screened achieved an average improvement of 2.32%, with a maximum of 6.6% in mAP75 on five different strong baseline detectors.

Accurate deep learning based method for real-time directly modulated laser modeling

Rate equations and numerical simulations relying on complex mathematical and physical principles are typically used to model directly modulated lasers (DMLs) but have difficulty simulating dynamic DML behavior in real-time under varying conditions due to their high complexity. Here, we introduce a data-driven deep learning method to model DMLs, aiming to achieve high accuracy with reduced computational complexity. This approach employs bidirectional long short-term memory (BiLSTM) enhanced by advanced feature recalibration and nonlinear fitting techniques. The result is compared with LSTM, standard BiLSTM, and recurrent neural network (RNN) architectures. The proposed model obtains the best results for the evaluated metrics. The satisfactory output waveforms and acceptable spectra indicate that the proposed model offers an accurate and real-time method to model DMLs.

Multistage deep learning for classification of Helicobacter pylori infection status using endoscopic images.

The automated classification of Helicobacter pylori infection status is gaining attention, distinguishing among uninfected (no history of H. pylori infection), current infection, and post-eradication. However, this classification has relatively low performance, primarily due to the intricate nature of the task. This study aims to develop a new multistage deep learning method for automatically classifying H. pylori infection status. The proposed multistage deep learning method was developed using a training set of 538 subjects, then tested on a validation set of 146 subjects. The classification performance of this new method was compared with the findings of four physicians. The accuracy of our method was 87.7%, 83.6%, and 95.9% for uninfected, post-eradication, and currently infected cases, respectively, whereas that of the physicians was 81.7%, 76.5%, and 90.3%, respectively. When including the patient's H. pylori eradication history information, the classification accuracy of the method was 92.5%, 91.1%, and 98.6% for uninfected, post-eradication, and currently infected cases, respectively, whereas that of the physicians was 85.6%, 85.1%, and 97.4%, respectively. The new multistage deep learning method shows potential for an innovative approach to gastric cancer screening. It can evaluate individual subjects' cancer risk based on endoscopic images and reduce the burden of physicians.

De novo designed proteins neutralize lethal snake venom toxins.

Snakebite envenoming remains a devastating and neglected tropical disease, claiming over 100,000 lives annually and causing severe complications and long-lasting disabilities for many more1,2. Three-finger toxins (3FTx) are highly toxic components of elapid snake venoms that can cause diverse pathologies, including severe tissue damage3 and inhibition of nicotinic acetylcholine receptors, resulting in life-threatening neurotoxicity4. At present, the only available treatments for snakebites consist of polyclonal antibodies derived from the plasma of immunized animals, which have high cost and limited efficacy against 3FTxs5-7. Here we used deep learning methods to de novo design proteins to bind short-chain and long-chain α-neurotoxins and cytotoxins from the 3FTx family. With limited experimental screening, we obtained protein designs with remarkable thermal stability, high binding affinity and near-atomic-level agreement with the computational models. The designed proteins effectively neutralized all three 3FTx subfamilies in vitro and protected mice from a lethal neurotoxin challenge. Such potent, stable and readily manufacturable toxin-neutralizing proteins could provide the basis for safer, cost-effective and widely accessible next-generation antivenom therapeutics. Beyond snakebite, our results highlight how computational design could help democratize therapeutic discovery, particularly in resource-limited settings, by substantially reducing costs and resource requirements for the development of therapies for neglected tropical diseases.

Deep Learning Method for Airfoil Flow Field Simulation Based on Unet++

ABSTRACTThis paper investigates the accuracy of U‐Net++ networks in predicting Reynolds‐Averaged Navier‐Stokes (RANS) solutions. The study employs the symbolic distance function (SDF) to represent geometry and flow conditions, utilizing parameterized airfoil data from the UIUC (University of Illinois at Urbana‐Champaign) airfoil datasets. The research assesses the performance of multiple trained neural networks in predicting pressure and velocity distributions. Specifically, the study examines the influence of varying network weights on solution accuracy. Through the optimization of the model, the research demonstrates that the mean relative error is below 1.72% for a range of previously unseen wing shapes, with a computational speedup factor of up to 1,000× in certain scenarios. The accuracy achieved by this model underscores the significant potential of deep learning‐based approaches as reliable tools for aerodynamic design and optimization.

NNFit: A Self-Supervised Deep Learning Method for Accelerated Quantification of High- Resolution Short Echo Time MR Spectroscopy Datasets.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To develop and evaluate the performance of NNFit, a self-supervised deep-learning method for quantification of high-resolution short echo-time (TE) echo-planar spectroscopic imaging (EPSI) datasets, with the goal of addressing the computational bottleneck of conventional spectral quantification methods in the clinical workflow. Materials and Methods This retrospective study included 89 short-TE whole-brain EPSI/GRAPPA scans from clinical trials for glioblastoma (Trial 1, May 2014-October 2018) and major-depressive-disorder (Trial 2, 2022- 2023). The training dataset included 685k spectra from 20 participants (60 scans) in Trial 1. The testing-dataset included 115k spectra from 5 participants (13 scans) in Trial 1 and 145k spectra from 7 participants (16 scans) in Trial 2. A comparative analysis was performed between NNFit and a widely used parametric-modeling spectral quantitation method (FITT). Metabolite maps generated by each method were compared using the structural- similarity-index-measure (SSIM) and linear-correlation-coefficient (R2). Radiation treatment volumes for glioblastoma based on the metabolite maps were compared with the Dice-coefficient and a two-tailed t test. Results Average SSIM and R2 scores for Trial 1 test set data were 0.91/0.90 (choline), 0.93/0.93 (creatine), 0.93/0.93 (n-acetylaspartate), 0.80/0.72 (myo-inositol), and 0.59/0.47 (glutamate + glutamine). Average scores for Trial 2 test set data were 0.95/0.95, 0.98/0.97, 0.98/0.98, 0.92/0.92, and 0.79/0.81 respectively. The treatment volumes had average Dice coefficient of 0.92. NNFit's average processing time was 90.1 seconds, whereas FITT took 52.9 minutes on average. Conclusion This study demonstrates that a deep learning approach to spectral quantitation offers comparable performance to conventional quantification methods for EPSI data, but with faster processing at short-TE. ©RSNA, 2025.

A Serial MRI-based Deep Learning Model to Predict Survival in Patients with Locoregionally Advanced Nasopharyngeal Carcinoma.

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To develop and evaluate a deep learning-based prognostic model for predicting survival in locoregionally- advanced nasopharyngeal carcinoma (LA-NPC) using serial MRI before and after induction chemotherapy (IC). Materials and Methods This multicenter retrospective study included 1039 LA-NPC patients (779 male, 260 female, mean age 44 [standard deviation: 11]) diagnosed between April 2009 and December 2015. A radiomics- clinical prognostic model (Model RC) was developed using pre-and post-IC MRI and other clinical factors using graph convolutional neural networks (GCN). The concordance index (C-index) was used to evaluate model performance in predicting disease-free survival (DFS). The survival benefits of concurrent chemoradiation therapy (CCRT) were analyzed in model-defined risk groups. Results The C-indexes of Model RC for predicting DFS were significantly higher than those of TNM staging in the internal (0.79 versus 0.53) and external (0.79 versus 0.62, both P < .001) testing cohorts. The 5-year DFS for the Model RC-defined low-risk group was significantly better than that of the high-risk group (90.6% versus 58.9%, P < .001). In high-risk patients, those who received CCRT had a higher 5-year DFS rate than those who did not (58.7% versus 28.6%, P = .03). There was no evidence of a difference in 5-year DFS rate in low-risk patients who did or did not receive CCRT (91.9% versus 81.3%, P = .19). Conclusion Serial MRI before and after IC can effectively predict survival in LA-NPC. The radiomics-clinical prognostic model developed using a GCN-based deep learning method showed good risk discrimination capabilities and may facilitate risk-adaptive therapy. ©RSNA, 2025.

Real-time detection of powder bed defects in laser powder bed fusion using deep learning on 3D point clouds

ABSTRACT Powder bed defects are critical factors affecting the print quality and stability in Laser Powder Bed Fusion (LPBF). However, traditional 2D image-based powder bed defect monitoring methods are limited by sensitivity to lighting conditions and insufficient data capture. This study proposes a real-time defect monitoring system based on 3D point cloud data and deep learning approach. The system uses binocular vision to capture point cloud data in real time, enabling high-precision defect segmentation with advanced deep learning models. However, direct deep learning on point clouds can result in the loss of small defect features during downsampling. To address this, an indirect point cloud deep learning method based on 2D projection is introduced, which improves segmentation accuracy for small defects while reducing inference time. By deploying the trained model, this study establishes a closed-loop control system for powder bed defect detection and conducts real-world printing tests, demonstrating effective defect remediation capabilities. Although larger-scale industrial testing is still required, this research illustrates the significant potential of 3D point cloud-based deep learning in enhancing defect detection and quality control in additive manufacturing.

Reverse design of broadband sound absorption structure based on deep learning method

This research presents a method based on deep learning for the reverse design of sound-absorbing structures. Traditional methods require time-consuming individual numerical simulations followed by cumbersome calculations, whereas the deep learning design method significantly simplifies the design process, achieving efficient and rapid design objectives. By utilizing deep neural networks, a mapping relationship between structural parameters and the sound absorption coefficient curve is established. The forward network predicts the sound absorption coefficient curve, while the reverse network enables the on-demand design of structural parameters for broadband high sound absorption. During the design process, a mean squared error (MSE) below 0.0001 is achieved. The accuracy of the proposed design method is validated through examples. The results demonstrate that the trained deep learning neural network could effectively replace the complex physical mechanisms between structural parameters and sound absorption coefficient curves. This deep learning design method could also be extended to other types of metamaterial reverse designs, significantly enhancing the efficiency of complex metamaterial designs. Lightweight design is crucial for energy saving and emission reduction. With the total mass and average sound absorption coefficient of sound-absorbing materials as targets, the NSGA-II algorithm has been used for multi-objective optimization design. The optimized average sound absorption coefficient increased by 4.84%, and the total material mass was reduced by 18.98%.