Deep Learning-based Methods Research Articles

DO-GMA: An End-to-End Drug-Target Interaction Identification Framework with a Depthwise Overparameterized Convolutional Network and the Gated Multihead Attention Mechanism.

Identification of potential drug-target interactions (DTIs) is a crucial step in drug discovery and repurposing. Although deep learning effectively deciphers DTIs, most deep learning-based methods represent drug features from only a single perspective. Moreover, the fusion method of drug and protein features needs further refinement. To address the above two problems, in this study, we develop a novel end-to-end framework named DO-GMA for potential DTI identification by incorporating Depthwise Overparameterized convolutional neural network and the Gated Multihead Attention mechanism with shared-learned queries and bilinear model concatenation. DO-GMA first designs a depthwise overparameterized convolutional neural network to learn drug representations from their SMILES strings and protein representations from their amino acid sequences. Next, it extracts drug representations from their 2D molecular graphs through a graph convolutional network. Subsequently, it fuses drug and protein features by combining the gated attention mechanism and the multihead attention mechanism with shared-learned queries and bilinear model concatenation. Finally, it takes the fused drug-target features as inputs and builds a multilayer perceptron to classify unlabeled drug-target pairs (DTPs). DO-GMA was benchmarked against six newest DTI prediction methods (CPI-GNN, BACPI, CPGL, DrugBAN, BINDTI, and FOTF-CPI) under four different experimental settings on four DTI data sets (i.e., DrugBank, BioSNAP, C.elegans, and BindingDB). The results show that DO-GMA significantly outperformed the above six methods based on AUC, AUPR, accuracy, F1-score, and MCC. An ablation study, robust statistical analysis, sensitivity analysis of parameters, visualization of the fused features, computational cost analysis, and case analysis further validated the powerful DTI identification performance of DO-GMA. In addition, DO-GMA predicted that two drug-protein pairs (i.e., DB00568 and P06276, and DB09118 and Q9UQD0) could be interacting. DO-GMA is freely available at https://github.com/plhhnu/DO-GMA.

Hybrid generative adversarial network based on frequency and spatial domain for histopathological image synthesis

BackgroundDue to the complexity and cost of preparing histopathological slides, deep learning-based methods have been developed to generate high-quality histological images. However, existing approaches primarily focus on spatial domain information, neglecting the periodic information in the frequency domain and the complementary relationship between the two domains. In this paper, we proposed a generative adversarial network that employs a cross-attention mechanism to extract and fuse features across spatial and frequency domains. The method optimizes frequency domain features using spatial domain guidance and refines spatial features with frequency domain information, preserving key details while eliminating redundancy to generate high-quality histological images.ResultsOur model incorporates a variable-window mixed attention module to dynamically adjust attention window sizes, capturing both local details and global context. A spectral filtering module enhances the extraction of repetitive textures and periodic structures, while a cross-attention fusion module dynamically weights features from both domains, focusing on the most critical information to produce realistic and detailed images.ConclusionsThe proposed method achieves efficient spatial-frequency domain fusion, significantly improving image generation quality. Experiments on the Patch Camelyon dataset show superior performance over eight state-of-the-art models across five metrics. This approach advances automated histopathological image generation with potential for clinical applications.

Transformer-Based Explainable Model for Breast Cancer Lesion Segmentation

Breast cancer is one of the most prevalent cancers among women, with early detection playing a critical role in improving survival rates. This study introduces a novel transformer-based explainable model for breast cancer lesion segmentation (TEBLS), aimed at enhancing the accuracy and interpretability of breast cancer lesion segmentation in medical imaging. TEBLS integrates a multi-scale information fusion approach with a hierarchical vision transformer, capturing both local and global features by leveraging the self-attention mechanism. This model addresses the limitations of existing segmentation methods, such as the inability to effectively capture long-range dependencies and fine-grained semantic information. Additionally, TEBLS incorporates visualization techniques to provide insights into the segmentation process, enhancing the model’s interpretability for clinical use. Experiments demonstrate that TEBLS outperforms traditional and existing deep learning-based methods in segmenting complex breast cancer lesions with variations in size, shape, and texture, achieving a mean DSC of 81.86% and a mean AUC of 97.72% on the CBIS-DDSM test set. Our model not only improves segmentation accuracy but also offers a more explainable framework, which has the potential to be used in clinical settings.

A Deep Learning-Based Method for Measuring Apparent Disease Areas of Sling Sheaths

The sling sheath plays an important protective role in the sling of suspension bridges, effectively preventing accidental damage to the sling caused by wind, fatigue and other impacts. To conduct a quantitative analysis of the apparent disease of suspension bridge slings, a method for segmenting and quantifying the apparent disease of the sling sheath using deep learning and image processing was proposed. A total of 1408 disease images were obtained after image acquisition of a suspension bridge following sling replacement. MATLAB 2021a Image Labeler software was used to establish a disease dataset by manual labelling. Then, the MobileNetV2 model was trained and tested on the dataset to determine disease segmentation; additionally, an area measurement algorithm was proposed based on the images’ projection relationships. Finally, the measurement results were compared with the manually acquired crack area. The results show that the accuracy of background and sheath category pixels in the MobileNetV2 model is above 97%, indicating that the model achieves satisfactory results in these classifications. However, the accuracy of crack category pixels and the intersection over union ratio only reaches 80%, which needs to be improved by setting model correction coefficients. When measuring directly, it was found that the area measurement error of the test image mainly ranged between 8% and 30%, and the measurement error of the crack area after correction mainly ranged between −3% and 15%, indicating that the area measurement method can achieve a higher degree of measurement accuracy. The method for segmenting and quantifying the apparent disease of the sling sheath based on deep learning and image processing fills the research gap in the measurement of the surface damage area caused by apparent disease and has the advantages of high efficiency and high recognition accuracy. Reducing the maintenance costs of suspension bridge slings is crucial for promoting comprehensive intelligent detection of bridges and advancing the smart transformation of the civil engineering industry.

Performance of Radiomics-based machine learning and deep learning-based methods in the prediction of tumor grade in meningioma: a systematic review and meta-analysis.

Currently, the World Health Organization (WHO) grade of meningiomas is determined based on the biopsy results. Therefore, accurate non-invasive preoperative grading could significantly improve treatment planning and patient outcomes. Considering recent advances in machine learning (ML) and deep learning (DL), this meta-analysis aimed to evaluate the performance of these models in predicting the WHO meningioma grade using imaging data. A systematic search was performed in PubMed/MEDLINE, Embase, and the Cochrane Library for studies published up to April 1, 2024, and reporting the performance metrics of the ML models in predicting of WHO meningioma grade using imaging studies. Pooled area under the receiver operating characteristics curve (AUROC), specificity, and sensitivity were estimated. Subgroup and meta-regression analyses were performed based on a number of potential influencing variables. A total of 32 studies with 15,365 patients were included in the present study. The overall pooled sensitivity, specificity, and AUROC of ML methods for prediction of tumor grade in meningioma were 85% (95% CI, 79-89%), 87% (95% CI, 81-91%), and 93% (95% CI, 90-95%), respectively. Both the type of validation and study cohort (training or test) were significantly associated with model performance. However, no significant association was found between the sample size or the type of ML method and model performance. The ML predictive models show a high overall performance in predicting the WHO meningioma grade using imaging data. Further studies on the performance of DL algorithms in larger datasets using external validation are needed.

Development of deep reinforcement learning-based fault diagnosis method for actuator faults in unmanned aerial vehicles

Abstract Actuator faults in unmanned aerial vehicles (UAVs) can have significant and potentially adverse effects on their safety and performance, highlighting the critical importance of fault diagnosis in UAV design. Ensuring the reliability of these systems in various applications often requires the use of advanced diagnostic algorithms. Artificial intelligence methods, such as deep learning and machine learning techniques, enable fault diagnosis through sample-based learning without the need for prior knowledge of fault mechanisms or physics-based models. However, UAV fault datasets are typically small due to stringent safety standards, which presents challenges for achieving high-performance fault diagnosis. To address this, deep reinforcement learning (DRL) algorithms offer a unique advantage by combining deep learning’s automatic feature extraction with reinforcement learning’s interactive learning approach, improving both learning capabilities and robustness. In this study, we propose and evaluate two DRL-based fault diagnosis models, which demonstrate remarkable accuracy in fault diagnosis, consistently exceeding $99{\rm{\% }}$ . Notably, under small sample scenarios, the proposed models significantly outperform traditional classifiers such as decision trees, support vector machines, and multilayer perceptron neural networks. These findings suggest that the integration of DRL enhances fault diagnosis performance, particularly in data-limited environments.

MAFF-Net: PCB defect detection via feature-enhancement fusion

Abstract Defect detection of printed circuit board (PCB) is of significant practical importance to ensure quality control in the production process. However, traditional defect detection methods suffer from limitations such as low detection accuracy and poor generalization ability. To tackle these issues, we propose a novel deep learning-based defect detection method for bare PCBs through multi-attention adaptive feature-enhancement fusion (AFF). First, we utilize ResNext101 as the backbone for feature extractor and embed a normalization-based attention mechanism in a residual structure, aiming at improving the feature extraction capability of the network. Second, we introduce an AFF module, which leverages multi-scale feature extraction and feature fusion to facilitate information interaction and enhance the correlation of feature information between channels. Finally, we incorporate the coordinate attention mechanism into AFF to highlight the target area for boosting detection accuracy. The experimental results demonstrate the effectiveness of the proposed method, which achieves a mean accuracy precision (mAP) of 99.01 % on a publicly available PCB defect dataset.

Retrospectively Quantified T2 Improves Detection of Clinically Significant Peripheral Zone Prostate Cancer

Background: Multiparametric MRI (mpMRI) as a non-invasive imaging tool is important in prostate cancer (PCa) detection and localization. Combined with radiomics analysis, features extracted from mpMRI have been utilized to predict PCa aggressiveness. T2 mapping provides quantitative information in PCa diagnoses but is not routinely available in clinical practice. Previous work from our group developed a deep learning-based method to estimate T2 maps from clinically acquired T1- and T2-weighted images. This study aims to evaluate the added value of the estimated T2 map by combining it with conventional T2-weighted images for detecting clinically significant PCa (csPCa). Methods: An amount of 76 peripheral zone prostate lesions, including clinically significant and insignificant cases, were retrospectively analyzed. Radiomic features were extracted from conventional T2-weighted images and deep learning-estimated T2 maps, followed by feature selection and model development using five-fold cross-validation. Logistic regression and Gaussian Process classifiers were employed to develop the prediction models, with performance evaluated by area under the curve (AUC) and accuracy metrics. Results: The model incorporating features from both T2-weighted images and estimated T2 maps achieved an AUC of 0.803, significantly outperforming the model based solely on T2-weighted image features (AUC of 0.700, p = 0.048). Conclusions: Radiomics features extracted from deep learning-estimated T2 maps provide additional quantitative information that improves the prediction of peripheral zone csPCa aggressiveness, potentially enhancing risk stratification in non-invasive PCa diagnostics.

Exploring Trends and Clusters in Human Posture Recognition Research: An Analysis Using CiteSpace

This study delves into interdisciplinary research directions in human posture recognition, covering vision-based and non-vision-based methods. Visually analyzing 3066 core research papers published from 2011 to 2024 with CiteSpace software reveals knowledge structures, research topics, key documents, trends, and institutional contributions. In-depth citation analysis identified 1200 articles and five significant research clusters. Findings show that in recent years, deep learning and sensor-based methods have dominated, significantly improving recognition accuracy, like the deep learning-based posture recognition method achieving 99.7% verification set accuracy with a 20-ms delay in a controlled environment. Logarithmic growth analysis of annual publications, supported by logistic model fitting, indicates the field’s maturation since 2011, with a shift from early simple applications of traditional and deep learning algorithms to integrating interdisciplinary approaches for problem-solving as the field matures and a predicted decline in future breakthroughs. By integrating indicators like citation bursts, degree centrality, and sigma, the research identifies interdisciplinary trends and key innovation directions, showing a transition from traditional to deep learning and multi-sensor data fusion methods. The integration of biomechanics principles with engineering technologies highlights new research paths. Overall, this study offers a systematic overview to identify gaps, trends, and innovation directions, facilitating future research and providing a roadmap for innovation in human posture recognition.

Syn2Real: synthesis of CT image ring artifacts for deep learning-based correction.

Objective.We strive to overcome the challenges posed by ring artifacts in X-ray computed tomography (CT) by developing a novel approach for generating training data for deep learning-based methods. Training such networks require large, high quality, datasets that are often generated in the data domain, time-consuming and expensive. Our objective is to develop a technique for synthesizing realistic ring artifacts directly in the image domain, enabling scalable production of training data without relying on specific imaging system physics.
Approach.We develop ''Syn2Real,'' a computationally efficient pipeline that generates realistic ring artifacts directly in the image domain. To demonstrate the effectiveness of our approach, we train two versions of UNet, vanilla and a high capacity version with self-attention layers that we call UNetpp, with ℓ2and perceptual losses, as well as a diffusion model, on energy-integrating CT images with and without these synthetic ring artifacts. 
Main Results.Despite being trained on conventional single-energy CT images, our models effectively correct ring artifacts across various monoenergetic images, at different energy levels and slice thicknesses, from a prototype photon-counting CT system. This generalizability validates the realism and versatility of our ring artifact generation process.
Significance.Ring artifacts in X-ray CT pose a unique challenge to image quality and clinical utility. By focusing on data generation, our work provides a foundation for developing more robust and adaptable ring artifact correction methods for pre-clinical, clinical and other CT applications.

Automated Quantification of Rebar Mesh Inspection in Hidden Engineering Structures via Deep Learning

This paper presents an in-depth study of the automated recognition and geometric information quantification of rebar meshes, proposing a deep learning-based method for rebar mesh detection and segmentation. By constructing a diverse rebar mesh image dataset, an improved Unet-based model was developed, incorporating residual modules to enhance the network’s feature extraction capabilities and training efficiency. The study found that the improved model maintains high segmentation accuracy and robustness even in the presence of complex backgrounds and noise. To achieve the precise measurement of rebar spacing, a rebar intersection detection algorithm based on convolution operations was designed, and the IQR (Interquartile Range) algorithm was applied to remove outliers, ensuring the accuracy and reliability of spacing calculations. The experimental results demonstrate that the proposed model and methods effectively and efficiently accomplish the automated recognition and geometric information extraction of rebar meshes, providing reliable technical support for the automated detection and geometric data analysis of rebar meshes in practical engineering applications.

Deciphering Protein Secondary Structures and Nucleic Acids in Cryo-EM Maps Using Deep Learning.

With the resolution revolution of cryo-electron microscopy (cryo-EM) and the rapid development of image processing technology, cryo-EM has become an indispensable experimental method for determining the three-dimensional structures of biological macromolecules. However, structural modeling from cryo-EM maps remains a difficult task for intermediate-resolution maps. In such cases, detection of protein secondary structures and nucleic acid locations in an EM map is of great value for model building of the map. Meeting the need, we present a deep learning-based method for detecting protein secondary structures and nucleic acid locations in cryo-EM density maps, named EMInfo. EMInfo was extensively evaluated on two protein-nucleic acid complex test sets including intermediate-resolution experimental maps and high-resolution experimental maps and compared them with two state-of-the-art methods including Emap2sec+ and Haruspex. It is shown that EMInfo can accurately predict different structural categories in an EM map. EMInfo is freely available at http://huanglab.phys.hust.edu.cn/EMInfo/.

Integrating pharmacogenomics and cheminformatics with diverse disease phenotypes for cell type-guided drug discovery

BackgroundLarge-scale pharmacogenomic resources, such as the Connectivity Map (CMap), have greatly assisted computational drug discovery. However, despite their widespread use, CMap-based methods have thus far been agnostic to the biological activity of drugs as well as to the genomic effects of drugs in multiple disease contexts. Here, we present a network-based statistical approach, Pathopticon, that uses CMap to build cell type-specific gene-drug perturbation networks and integrates these networks with cheminformatic data and diverse disease phenotypes to prioritize drugs in a cell type-dependent manner.MethodsWe build cell type-specific gene-drug perturbation networks from CMap data using a statistical procedure we call Quantile-based Instance Z-score Consensus (QUIZ-C). Using these networks and a large-scale disease-gene network consisting of 569 disease signatures from the Enrichr database, we calculate Pathophenotypic Congruity Scores (PACOS) between input gene signatures and drug perturbation signatures and combine these scores with cheminformatic data from ChEMBL to prioritize drugs. We benchmark our approach by calculating area under the receiver operating characteristic curves (AUROC) for 73 gene sets from the Molecular Signatures Database (MSigDB) using target gene expression profiles from the Comparative Toxicogenomics Database (CTD). We validate the drugs predicted in our proofs-of-concept using real-time polymerase chain reaction (qPCR) experiments.ResultsCell type-specific gene-drug perturbation networks built using QUIZ-C are topologically distinct, reflecting the biological uniqueness of the cell lines in CMap, and are enriched in known drug targets. Pathopticon demonstrates a better prediction performance than solely cheminformatic measures as well as state-of-the-art network and deep learning-based methods. Top predictions made by Pathopticon have high chemical structural diversity, suggesting their potential for building compound libraries. In proof-of-concept applications on vascular diseases, we demonstrate that Pathopticon helps guide in vitro experiments by identifying pathways that are potentially regulated by the predicted therapeutic candidates.ConclusionsOur network-based analytical framework integrating pharmacogenomics and cheminformatics (available at https://github.com/r-duh/Pathopticon) provides a feasible blueprint for a cell type-specific drug discovery and repositioning platform with broad implications for the efficiency and success of drug development.

Fast 3D Transmission Tower Detection Based on Virtual Views

Advanced remote sensing technologies leverage extensive synthetic aperture radar (SAR) satellite data and high-resolution airborne light detection and ranging (LiDAR) data to swiftly capture comprehensive 3D information about electrical grid assets and their surrounding environments. This facilitates in-depth scene analysis for target detection and classification, allowing for the early recognition of potential hazards near transmission towers (TTs). These innovations present a groundbreaking strategy for the automated inspection of electrical grid assets. However, traditional 3D target detection techniques, which involve searching the entire 3D space, are marred by low accuracy and high computational demands. Although deep learning-based 3D target detection methods have significantly improved detection precision, they rely on a large volume of 3D target samples for training and are sensitive to point cloud data density. Moreover, these methods demonstrate low detection efficiency, constraining their application in the automated monitoring of electricity networks. This paper proposes a fast 3D target detection method using virtual views to overcome these challenges related to detection accuracy and efficiency. The method first utilizes cutting-edge 2D splatting technology to project 3D point clouds with diverse densities from a specific viewpoint, generating a 2D virtual image. Then, a novel local–global dual-path feature fusion network based on YOLO is applied to detect TTs on the virtual image, ensuring efficient and accurate identification of their positions and types. Finally, by leveraging the projection transformation between the virtual image and the 3D point cloud, combined with a 3D region growing algorithm, the 3D points belonging to the TTs are extracted from the whole 3D point cloud. The effectiveness of the proposed method in terms of target detection rate and efficiency is validated through experiments on synthetic datasets and outdoor LiDAR point clouds.

Multi-Target Detection Method of Intelligent Driving Traffic Scene Based on Faster R-CNN++

The current multi-objective detection methods are plagued by several issues, including insufficient detection efficiency, slow speeds, and difficulties in deciphering driving rules. To address these challenges, we introduce a traffic scene classification method that utilizes Improved Faster Regions with Convolutional Neural Network features (Faster R-CNN++). Initially, the shared convolutional layer and the Recurrent Criss-Cross Attention (RCCA) layer are employed to extract features from the input image. Subsequently, the derived feature map is fed into the Region Proposal Network (RPN) to generate detection boxes, and the Region of Interest Align (RoI Align) layer selects features corresponding to each Region of Interest (RoI) on the feature map, guided by the RPN’s output. Ultimately, we adopt an alternating training approach. Simulation experiments confirm the effectiveness of our method in multi-scene object detection. Our method has demonstrated significant improvements over existing algorithms and has delivered outstanding performance on the BDD-100[Formula: see text]K, ApolloScape, and NuScenes datasets. The results indicate that our deep learning-based traffic scene classification method can accurately discern the behavioral characteristics of various traffic participants.

A diesel engine valve leakage fault diagnostic method using modified VGG19 and transfer learning

ABSTRACT Traditional deep learning-based fault diagnosis methods for diesel engine valve leakage, often rely on large datasets. This study proposes the GAP-VGG19-BN method based on VGG19 and transfer learning. Global average pooling (GAP) is adopted to replace the full connection layer and improve parameter reduction. Batch Normalization (BN) layers are set after each convolution layer to enhance the computational speed. We build a small-scale dataset using the time–frequency and the wavelet methods constructed from the diesel engine cylinder head vibration signal. The GAP-VGG19-BN is pre-trained through ImageNet; then, the well-trained parameters are transferred onto the constructed dataset. The BN method is adopted to make the learning rate approach obtain an optimal global solution of learning rate. Experimental results show that the proposed method has higher accuracy, stronger robustness, less computational cost, and better generalisation for small-scale datasets. The proposed method enriches the fault diagnosis method of diesel engine valve leakage.

Efficient Personalization in E-Commerce: Leveraging Universal Customer Representations with Embeddings

E-commerce has grown into a billion-dollar industry in recent times with an ever-increasing number of individuals using it regularly. Thus, e-commerce companies can gather interaction data from their customers and analyze it to create focused and personalized marketing campaigns. For large companies, it is possible to tap into these data for personalization using deep learning-based methods that require enormous computing resources. Small companies, on the other hand, cannot afford this. Furthermore, this level of tailor-made addressing necessitates an accurate customer representation. Nevertheless, the exploration of universal representations applicable across various tasks has been limited despite the advantages they offer. We propose a universal customer representation learned only from customer interaction data. We demonstrate that self-supervised trained embeddings of the customer interaction context are a suitable universal customer representation for various e-commerce tasks. To demonstrate the effectiveness of our approach, we conducted experiments comparing four different state-of-the-art approaches and their capabilities in different prediction tasks. Not only do we show that our method outperforms others in most cases, but it also meets other important criteria for real-world applications. It is particularly important to emphasize that our approach does not require a significant amount of resources, and furthermore, is data protection compliant by not using personal information.

The segmentation of nanoparticles with a novel approach of HRU2-Net†

Nanoparticles have great potential for the application in new energy and aerospace fields. The distribution of nanoparticle sizes is a critical determinant of material properties and serves as a significant parameter in defining the characteristics of zero-dimensional nanomaterials. In this study, we proposed HRU2-Net†, an enhancement of the U2-Net† model, featuring multi-level semantic information fusion. This approach exhibits strong competitiveness and refined segmentation capabilities for nanoparticle segmentation. It achieves a Mean intersection over union (MIoU) of 87.31%, with an accuracy rate exceeding 97.31%, leading to a significant improvement in segmentation effectiveness and precision. The results show that the deep learning-based method significantly enhances the efficacy of nanomaterial research, which holds substantial significance for the advancement of nanomaterial science.

A spatial-temporal hierarchical modeling framework for multi-step ride-hailing demand forecasting

Forecasting urban ride-hailing demand is the main duty in Intelligent Transportation Systems. Accurate predicting can significantly improve the efficiency of regional capacity allocation, promoting energy conversation and emission reduction. Currently, the mainstream approaches utilize advanced deep learning-based methods to model aspects of both temporal and spatial. However, these previous methods overlook the long-term temporal dependence of selection and the hierarchical nature of urban road networks. To this end, we propose a novel Spatial-Temporal Hierarchical Network (STHNet) for urban ride-hailing demand prediction. Specially, The Depthwise Separable Convolution Nerual Networks (DSCNNs) extract spatial features of the road network through channel-wise and point-wise convolution operations, while the Nested Long Short-Term Memory networks (NLSTMs) are used to capture the hierarchical temporal dependencies in sequential data. DSCNNs and NLSTMs are cascaded to form the basic module of the multi-step prediction framework, called Spatial-Temporal Hierarchical block. The block can be easily extended to other spatial-temporal modeling tasks. At the end of the network, we introduce a 3DCNN to learn Spatial-Temporal heterogeneity and integrate information. Furthermore, Teacher Forcing and secondary information are incorporated into STHNet to enhance training efficiency. Extensive experiment are conducted on a Xiamen transportation network with 64 regions shows that STHNet outperforms multiple State-Of-The-Art baselines. The qualitative results underscoring the practical engineering applicability of the proposed model.

Deep learning-based polarization 3D imaging method for underwater targets.

The significant absorption and scattering of light during its propagation in water severely degrade the quality of underwater imaging, presenting challenges for developing high-precision 3D imaging techniques based on optical methods. Polarization imaging has demonstrated effectiveness in mitigating the effects of scattering, making it a valuable approach for underwater imaging. Additionally, the polarization state of reflected light can be utilized for surface normal estimation and 3D shape reconstruction. This paper presents a learning-based method for 3D shape reconstruction of underwater targets using shape from polarization techniques. To address the lack of publicly available datasets for underwater polarization 3D imaging, we have developed a data acquisition system that simulates Jerlov Type I water conditions, creating a dataset of underwater polarized images along with corresponding ground truth surface normal images. Furthermore, we propose a network framework based on Attention U2Net for the 3D reconstruction of underwater polarized images. This framework is designed to capture detailed texture information of underwater targets and incorporates an effective polarization representation to resolve azimuthal ambiguity, thus enhancing the accuracy of underwater 3D imaging. Experimental results demonstrate that our method effectively addresses azimuthal ambiguity, reduces texture loss during reconstruction, and improves the accuracy of surface normal estimation, achieving superior performance compared to existing methods.