Cancer. 2025 Jun 1;131(11):e35883. doi: 10.1002/cncr.35883.

NO ABSTRACT

PMID:40457864 | DOI:10.1002/cncr.35883

J Phys Chem A. 2025 Jun 2. doi: 10.1021/acs.jpca.5c01881. Online ahead of print.

ABSTRACT

The aqueous diffusion coefficients of organic compounds are among the most important topics of interest for chemical research and engineering, particularly for the dispersal of organic pollutants in water environments. The experimental determination of the diffusion coefficients is time-consuming, and the experimental data are lacking for most newly synthesized chemicals. Our study developed a multimodal deep learning model incorporating molecular images, molecular descriptors, and temperatures to predict aqueous diffusion coefficients at varying temperatures. The deep learning model made accurate predictions on the test set (R2 = 0.986) and had a considerably lower error than previous empirical equations. Besides, the model interpretation indicated that the deep learning model correctly captured the effects of molecular features and temperature on the prediction of diffusion coefficients. Thus, our study provided a valuable tool for the rapid and accurate prediction of organic compounds diffusion coefficients in water at varying temperatures and also facilitated a better understanding of how different molecular features and water temperatures influence diffusion-controlled processes in both the environmental and engineered systems.

PMID:40457760 | DOI:10.1021/acs.jpca.5c01881

Magn Reson Med. 2025 Jun 2. doi: 10.1002/mrm.30591. Online ahead of print.

ABSTRACT

PURPOSE: To examine the effect of incorporating self-supervised denoising as a pre-processing step for training deep learning (DL) based reconstruction methods on data corrupted by Gaussian noise. K-space data employed for training are typically multi-coil and inherently noisy. Although DL-based reconstruction methods trained on fully sampled data can enable high reconstruction quality, obtaining large, noise-free datasets is impractical.

METHODS: We leverage Generalized Stein's Unbiased Risk Estimate (GSURE) for denoising. We evaluate two DL-based reconstruction methods: Diffusion Probabilistic Models (DPMs) and Model-Based Deep Learning (MoDL). We evaluate the impact of denoising on the performance of these DL-based methods in solving accelerated multi-coil magnetic resonance imaging (MRI) reconstruction. The experiments were carried out on T2-weighted brain and fat-suppressed proton-density knee scans.

RESULTS: We observed that self-supervised denoising enhances the quality and efficiency of MRI reconstructions across various scenarios. Specifically, employing denoised images rather than noisy counterparts when training DL networks results in lower normalized root mean squared error (NRMSE), higher structural similarity index measure (SSIM) and peak signal-to-noise ratio (PSNR) across different SNR levels, including 32, 22, and 12 dB for T2-weighted brain data, and 24, 14, and 4 dB for fat-suppressed knee data.

CONCLUSION: We showed that denoising is an essential pre-processing technique capable of improving the efficacy of DL-based MRI reconstruction methods under diverse conditions. By refining the quality of input data, denoising enables training more effective DL networks, potentially bypassing the need for noise-free reference MRI scans.

PMID:40457510 | DOI:10.1002/mrm.30591

Mol Cancer. 2025 Jun 2;24(1):159. doi: 10.1186/s12943-025-02369-9.

ABSTRACT

Cancer continues to be a significant international health issue, which demands the invention of new methods for early detection, precise diagnoses, and personalized treatments. Artificial intelligence (AI) has rapidly become a groundbreaking component in the modern era of oncology, offering sophisticated tools across the range of cancer care. In this review, we performed a systematic survey of the current status of AI technologies used for cancer diagnoses and therapeutic approaches. We discuss AI-facilitated imaging diagnostics using a range of modalities such as computed tomography, magnetic resonance imaging, positron emission tomography, ultrasound, and digital pathology, highlighting the growing role of deep learning in detecting early-stage cancers. We also explore applications of AI in genomics and biomarker discovery, liquid biopsies, and non-invasive diagnoses. In therapeutic interventions, AI-based clinical decision support systems, individualized treatment planning, and AI-facilitated drug discovery are transforming precision cancer therapies. The review also evaluates the effects of AI on radiation therapy, robotic surgery, and patient management, including survival predictions, remote monitoring, and AI-facilitated clinical trials. Finally, we discuss important challenges such as data privacy, interpretability, and regulatory issues, and recommend future directions that involve the use of federated learning, synthetic biology, and quantum-boosted AI. This review highlights the groundbreaking potential of AI to revolutionize cancer care by making diagnostics, treatments, and patient management more precise, efficient, and personalized.

PMID:40457408 | DOI:10.1186/s12943-025-02369-9

BMC Bioinformatics. 2025 Jun 2;26(1):148. doi: 10.1186/s12859-025-06173-6.

ABSTRACT

BACKGROUND: Understanding cellular heterogeneity within tissues hinges on knowledge of their spatial context. However, it is still challenging to accurately map cells to their spatial coordinates.

RESULTS: We present SC2Spa, a deep learning-based approach that learns intricate spatial relationships from spatial transcriptomics (ST) data. Benchmarking tests show that SC2Spa outperformed other predictors and accurately detected tissue architecture from transcriptome. SC2Spa successfully mapped single cell RNA sequencing (scRNA-seq) to Visium assay, providing an approach to enhance the resolution for low resolution ST data. Our test showed that SC2Spa performs well for various ST technologies and robust to spatial resolution. In addition, SC2Spa can suggest spatially variable genes that cannot be identified from previous approaches.

CONCLUSIONS: SC2Spa is a robust and accurate approach to provide single cells with their spatial location and identify spatially meaningful genes.

PMID:40457183 | DOI:10.1186/s12859-025-06173-6

NPJ Digit Med. 2025 Jun 2;8(1):293. doi: 10.1038/s41746-025-01571-9.

ABSTRACT

We employed a three-phase approach, culminating in a randomized controlled trial, to assess the efficacy of 3D-printed liver models in hepatobiliary surgical planning. Phase one involved developing and selecting 35 optimal 3DP models based on timeliness, cost, precision, and alignment with digital simulations. Phase two utilized deep learning algorithms to optimize the 3D reconstruction process, significantly enhancing efficiency and accuracy compared to manual segmentation. In phase three, a randomized controlled trial with 64 patients compared surgical outcomes between those planned with AI-enhanced physical 3DP models and those with traditional digital simulations. Results demonstrated that 3DP models were produced rapidly (3.52 h at $152 each) with high precision, AI-assisted reconstruction reduced processing time (303.5 vs. 557 min), and patients using AI-enhanced physical 3DP models experienced less intraoperative blood loss. Integrating deep learning with 3D printing offers a cost-effective, scalable method to enhance surgical planning and outcomes in hepatobiliary surgery.

PMID:40457016 | DOI:10.1038/s41746-025-01571-9

J Imaging Inform Med. 2025 Jun 2. doi: 10.1007/s10278-025-01559-7. Online ahead of print.

ABSTRACT

This study addresses the critical challenge of detecting out-of-distribution (OOD) chest X-rays, where subtle view differences between lateral and frontal radiographs can lead to diagnostic errors. We develop a GAN-based framework that learns the inherent feature distribution of frontal views from the MIMIC-CXR dataset through latent space optimization and Kolmogorov-Smirnov statistical testing. Our approach generates similarity scores to reliably identify OOD cases, achieving exceptional performance with 100% precision, and 97.5% accuracy in detecting lateral views. The method demonstrates consistent reliability across operating conditions, maintaining accuracy above 92.5% and precision exceeding 93% under varying detection thresholds. These results provide both theoretical insights and practical solutions for OOD detection in medical imaging, demonstrating how GANs can establish feature representations for identifying distributional shifts. By significantly improving model reliability when encountering view-based anomalies, our framework enhances the clinical applicability of deep learning systems, ultimately contributing to improved diagnostic safety and patient outcomes.

PMID:40457001 | DOI:10.1007/s10278-025-01559-7

J Imaging Inform Med. 2025 Jun 2. doi: 10.1007/s10278-025-01564-w. Online ahead of print.

ABSTRACT

Scoliosis is a three-dimensional spinal deformity that is the most common among spinal deformities and causes extremely serious posture disorders in advanced stages. Scoliosis can lead to various health problems, including pain, respiratory dysfunction, heart problems, mental health disorders, stress, and emotional difficulties. The current gold standard for grading scoliosis and planning treatment is based on the Cobb angle measurement on X-rays. The Cobb angle measurement is performed by physical medicine and rehabilitation specialists, orthopedists, radiologists, etc., in branches dealing with the musculoskeletal system. Manual calculation of the Cobb angle for this process is subjective and takes more time. Deep learning-based systems that can evaluate the Cobb angle objectively have been frequently used recently. In this article, we propose an enhanced ViT that allows doctors to evaluate the diagnosis of scoliosis more objectively without wasting time. The proposed model uses a custom attention mechanism instead of the standard multi-head attention mechanism for the ViT model. A dataset with 7 different classes was obtained from 1456 patients in total from Elazığ Fethi Sekin City Hospital Physical Medicine and Rehabilitation Clinic. Multiple models were used to compare the proposed architecture in the classification of scoliosis disease. The proposed improved ViT architecture exhibited the best performance with 95.21% accuracy. This result shows that a superior classification success was achieved compared to ResNet50, Swin Transformer, and standard ViT models.

PMID:40457000 | DOI:10.1007/s10278-025-01564-w

J Imaging Inform Med. 2025 Jun 2. doi: 10.1007/s10278-025-01553-z. Online ahead of print.

ABSTRACT

Distinguishing benign from malignant vertebral compression fractures is critical for clinical management but remains challenging on contrast-enhanced abdominal CT, which lacks the soft tissue contrast of MRI. This study evaluates and compares radiomic feature-based machine learning and convolutional neural network-based deep learning models for classifying VCFs using abdominal CT. A retrospective cohort of 447 vertebral compression fractures (196 benign, 251 malignant) from 286 patients was analyzed. Radiomic features were extracted using PyRadiomics, with Recursive Feature Elimination selecting six key texture-based features (e.g., Run Variance, Dependence Non-Uniformity Normalized), highlighting textural heterogeneity as a malignancy marker. Machine learning models (XGBoost, SVM, KNN, Random Forest) and a 3D CNN were trained on CT data, with performance assessed via precision, recall, F1 score, accuracy, and AUC. The deep learning model achieved marginally superior overall performance, with a statistically significant higher AUC (77.66% vs. 75.91%, p < 0.05) and better precision, F1 score, and accuracy compared to the top-performing machine learning model (XGBoost). Deep learning's attention maps localized diagnostically relevant regions, mimicking radiologists' focus, whereas radiomics lacked spatial interpretability despite offering quantifiable biomarkers. This study underscores the complementary strengths of machine learning and deep learning: radiomics provides interpretable features tied to tumor heterogeneity, while DL autonomously extracts high-dimensional patterns with spatial explainability. Integrating both approaches could enhance diagnostic accuracy and clinician trust in abdominal CT-based VCF assessment. Limitations include retrospective single-center data and potential selection bias. Future multi-center studies with diverse protocols and histopathological validation are warranted to generalize these findings.

PMID:40456998 | DOI:10.1007/s10278-025-01553-z

Sci Rep. 2025 Jun 2;15(1):19298. doi: 10.1038/s41598-025-04503-5.

ABSTRACT

Retinal disorders, posing significant risks of the loss of vision or blindness, are increasingly prevalent, due to factors such as the aging population and chronic conditions like diabetes. Traditional diagnostic methods, relying on manually analyzing images, often have problems making an early detection and with their accuracy and efficiency, largely due to the subjectivity of human judgment and the time-consuming nature of the process. This study introduces a novel AI-based framework for diagnosing retinal disease, referred to as RetinaDNet. This framework leverages dual-branch input, incorporating both retinal images and vessel segmentation images, along with transfer learning and ensemble learning algorithms. This enhances the accuracy of the diagnoses and the stability of the model, particularly in scenarios with small sample sizes. By using vascular features and mitigating the risk of overfitting, this framework demonstrates superior performance in terms of multiple metrics. In particular, a soft voting classifier combined with the ResNet50 model attains accuracy rate of 99.2% on the diabetic retinopathy diagnosis task, and 98.8% on the retina disease classification task. The source code can be accessed at https://github.com/yu0809/Dual-branch-retinal-diseases .

PMID:40456971 | DOI:10.1038/s41598-025-04503-5

Sci Rep. 2025 Jun 2;15(1):19274. doi: 10.1038/s41598-025-03766-2.

ABSTRACT

Hand recognition, the process of identifying or characterizing human hands in images or video streams, plays significant role in the biometrics, robotics, computer vision, and human-computer interaction. This technology relies on analyzing hand attributes such as shape, size, color, texture, and motion to perform tasks as gesture identification, hand tracking, and sign language interpretation. In particular, hand movement decoding from electromyography (EMG) signals has shown promise for understanding neuromuscular function and aiding in diagnosis and therapy for neuromuscular issues. Existing approaches range from deep learning techniques such as Long Short-Term Memory (LSTM) and Bidirectional LSTM (Bi-LSTM) to conventional machine learning methods like Support Vector Machines (SVM) and Random Forest. Deep learning automates the process, reducing the dependency on manual feature extraction. However, the performance of these models is heavily influenced by hyperparameters such as the number of neurons, hidden layers, and learning rates. This study proposes a novel method that uses the Grey Wolf Optimizer (GWO) to fine-tune the hyperparameters of a Bi-LSTM-based EMG classification system. Implemented in MATLAB R2021a, this approach aims to enhance the accuracy of Bi-LSTM models in categorizing EMG signals. Performance metrics such as accuracy of 95%, precision of 93%, F1-score of 94%, and recall of 91% are used for thorough evaluation. By leveraging GWO for hyperparameter optimization, the study aims to achieve more accurate diagnosis and efficient tracking of rehabilitation outcomes for patients with neuromuscular disorders. This research demonstrates the potential of integrating biomedical engineering and computational intelligence to empower individuals with neuromuscular disabilities, thereby enhancing their quality of life.

PMID:40456840 | DOI:10.1038/s41598-025-03766-2

Sci Rep. 2025 Jun 2;15(1):19309. doi: 10.1038/s41598-025-02649-w.

ABSTRACT

The rapid evolution of smart grids, driven by rising global energy demand and renewable energy integration, calls for intelligent, adaptive, and energy-efficient resource allocation strategies. Traditional energy management methods, based on static models or heuristic algorithms, often fail to handle real-time grid dynamics, leading to suboptimal energy distribution, high operational costs, and significant energy wastage. To overcome these challenges, this paper presents ORA-DL (Optimized Resource Allocation using Deep Learning) an advanced framework that integrates deep learning, Internet of Things (IoT)-based sensing, and real-time adaptive control to optimize smart grid energy management. ORA-DL employs deep neural networks, reinforcement learning, and multi-agent decision-making to accurately predict energy demand, allocate resources efficiently, and enhance grid stability. The framework leverages both historical and real-time data for proactive power flow management, while IoT-enabled sensors ensure continuous monitoring and low-latency response through edge and cloud computing infrastructure. Experimental results validate the effectiveness of ORA-DL, achieving 93.38% energy demand prediction accuracy, improving grid stability to 96.25%, and reducing energy wastage to 12.96%. Furthermore, ORA-DL enhances resource distribution efficiency by 15.22% and reduces operational costs by 22.96%, significantly outperforming conventional techniques. These performance gains are driven by real-time analytics, predictive modelling, and adaptive resource modulation. By combining AI-driven decision-making, IoT sensing, and adaptive learning, ORA-DL establishes a scalable, resilient, and sustainable energy management solution. The framework also provides a foundation for future advancements, including integration with edge computing, cybersecurity measures, and reinforcement learning enhancements, marking a significant step forward in smart grid optimization.

PMID:40456783 | DOI:10.1038/s41598-025-02649-w

IEEE Trans Med Imaging. 2025 Jun 2;PP. doi: 10.1109/TMI.2025.3575402. Online ahead of print.

ABSTRACT

Deep learning-based disease grading technologies facilitate timely medical intervention due to their high efficiency and accuracy. Recent advancements have enhanced grading performance by incorporating the ordinal relationships of disease labels. However, existing methods often assume same probability distributions for disease labels across instances within the same category, overlooking variations in label distributions. Additionally, the hyperparameters of these distributions are typically determined empirically, which may not accurately reflect the true distribution. To address these limitations, we propose a disease grading network utilizing a sample-aware asymmetric Gaussian label distribution, termed DGN-AGLD. This approach includes a variance predictor designed to learn and predict parameters that control the asymmetry of the Gaussian distribution, enabling distinct label distributions within the same category. This module can be seamlessly integrated into standard deep learning networks. Experimental results on four disease datasets validate the effectiveness and superiority of the proposed method, particularly on the IDRiD dataset, where it achieves a diabetic retinopathy accuracy of 77.67%. Furthermore, our method extends to joint disease grading tasks, yielding superior results and demonstrating significant generalization capabilities. Visual analysis indicates that our method more accurately captures the trend of disease progression by leveraging the asymmetry in label distribution. Our code is publicly available on https://github.com/ahtwq/AGNet.

PMID:40456095 | DOI:10.1109/TMI.2025.3575402

IEEE Trans Neural Netw Learn Syst. 2025 Jun 2;PP. doi: 10.1109/TNNLS.2025.3566647. Online ahead of print.

ABSTRACT

Visual recognition in low-light environments is a challenging problem since degraded images are the stacking of multiple degradations (noise, low light and blur, etc.). It has received extensive attention from academia and industry in the era of deep learning. Existing surveys focus on low-light image enhancement (LLIE) methods and normal-light visual recognition methods, while few comprehensive surveys of low-light-related vision tasks. This article provides a comprehensive survey of the latest advancements in low-light vision, including methods, datasets, and evaluation metrics, in two aspects: visual quality-driven and recognition quality-driven. On the visual quality-driven aspect, we survey a large number of very recent LLIE methods. On the recognition quality-driven aspect, we survey low-light object detection techniques in the deep learning era using more intuitive categorization method. Furthermore, a quantitative benchmarking of different methods is conducted on several widely adopted low-light vision-related datasets. Finally, we discuss the challenges that exist in low-light vision and future directions worth exploring. We provide a public website that will continue to track developments in this promising field.

PMID:40456083 | DOI:10.1109/TNNLS.2025.3566647

J Chem Inf Model. 2025 Jun 2. doi: 10.1021/acs.jcim.5c00584. Online ahead of print.

ABSTRACT

In the realm of medicinal chemistry, the predominant challenge in molecular design lies in managing extensive molecular data sets and effectively screening for, as well as preserving, molecules with potential value. Traditional methodologies typically utilize deep learning models or genetic algorithms (GA) for optimization, yet each approach has inherent limitations: deep learning models are constrained by substantial computational resource demands; genetic algorithms often yield molecular structures with low validity and feasibility. To overcome these challenges, we have developed the Molecular multiobjective optimization of Monte Carlo Tree Search (MCTS) and Non-Superiority Ranking Genetic Algorithm II (NSGA-II)-MNopt, which ingeniously integrates MCTS with NSGA-II. Specifically, NSGA-II demonstrates unique strengths in balancing multiple optimization objectives and achieves rapid performance through its crowding distance and nondominated ordering mechanisms, while MCTS focuses on enhancing the validity of molecular structures to ensure that the generated molecules are both desirable and feasible. Notably, MNopt does not require reliance on extensive molecular training data sets in the initial stages, effectively mitigating excessive resource consumption. Experimental results demonstrate that MNopt surpasses existing techniques in multiobjective optimization, generating effective and diverse molecular structures, thereby offering a crucial tool for novel drug discovery and materials science.

PMID:40456025 | DOI:10.1021/acs.jcim.5c00584

PLoS One. 2025 Jun 2;20(6):e0320701. doi: 10.1371/journal.pone.0320701. eCollection 2025.

ABSTRACT

Advancements in deep learning have revolutionized numerous real-world applications, including image recognition, visual question answering, and image captioning. Among these, image captioning has emerged as a critical area of research, with substantial progress achieved in Arabic, Chinese, Uyghur, Hindi, and predominantly English. However, despite Urdu being a morphologically rich and widely spoken language, research in Urdu image captioning remains underexplored due to a lack of resources. This study creates a new Urdu Image Captioning Dataset (UCID) called UC-23-RY to fill in the gaps in Urdu image captioning. The Flickr30k dataset inspired the 159,816 Urdu captions in the dataset. Additionally, it suggests deep learning architectures designed especially for Urdu image captioning, including NASNetLarge-LSTM and ResNet-50-LSTM. The NASNetLarge-LSTM and ResNet-50-LSTM models achieved notable BLEU-1 scores of 0.86 and 0.84 respectively, as demonstrated through evaluation in this study accessing the model's impact on caption quality. Additionally, it provides useful datasets and shows how well-suited sophisticated deep learning models are for improving automatic Urdu image captioning.

PMID:40455832 | DOI:10.1371/journal.pone.0320701

PLoS One. 2025 Jun 2;20(6):e0320732. doi: 10.1371/journal.pone.0320732. eCollection 2025.

ABSTRACT

BACKGROUND: Breast cancer is the most common malignant tumor among women worldwide, and early diagnosis is crucial for reducing mortality rates. Traditional diagnostic methods have significant limitations in terms of accuracy and consistency. Imaging is a common technique for diagnosing and predicting breast cancer, but human error remains a concern. Increasingly, artificial intelligence (AI) is being employed to assist physicians in reducing diagnostic errors.

METHODS: We developed an intelligent diagnostic model combining deep learning and radiomics to enhance breast tumor diagnosis. The model integrates MobileNet with ResNeXt-inspired depthwise separable and grouped convolutions, improving feature processing and efficiency while reducing parameters. Using AI-Dhabyani and TCIA breast ultrasound datasets, we validated the model internally and externally, comparing it to VGG16, ResNet, AlexNet, and MobileNet. Results: The internal validation set achieved an accuracy of 83.84% with an AUC of 0.92, outperforming other models. The external validation set showed an accuracy of 69.44% with an AUC of 0.75, demonstrating high robustness and generalizability. Conclusions: We developed an intelligent diagnostic model using deep learning and radiomics to improve breast tumor diagnosis. The model combines MobileNet with ResNeXt-inspired depthwise separable and grouped convolutions, enhancing feature processing and efficiency while reducing parameters. It was validated internally and externally using the AI-Dhabyani and TCIA breast ultrasound datasets and compared with VGG16, ResNet, AlexNet, and MobileNet.

PMID:40455816 | DOI:10.1371/journal.pone.0320732

PLoS One. 2025 Jun 2;20(6):e0323440. doi: 10.1371/journal.pone.0323440. eCollection 2025.

ABSTRACT

Intestinal sounds, primarily generated by the movement of digested gas and liquids during peristalsis, are acoustic signals that provide valuable insights into intestinal functioning. Traditionally, doctors have relied on stethoscopes to assess the degree of gastrointestinal activity. Recent advancements in computer-aided technologies and electronic stethoscopes have enhanced the understanding and analysis of these sounds. Studies utilizing advanced techniques like deep learning and convolutional neural networks have shown promise in analyzing bowel sounds. Nevertheless, the reliance on personal judgment and the need for large labeled datasets limit the broader applicability of these methods. This study introduces an innovative, unsupervised grading system to objectively evaluate gastrointestinal motility by analyzing bowel sounds through spectral feature analysis. This system offers a practical alternative to traditional listening techniques or complex models. It computes an activity score for digital audio using a cost-effective numerical grading method to assist doctors in quantifying gastrointestinal motility. The method's reliability, validated by Spearman's rank correlation, confirms its accuracy in assessing activity levels and highlights its potential as a reliable and practical tool for supporting objective medical assessments of bowel activity.

PMID:40455773 | DOI:10.1371/journal.pone.0323440

PLoS One. 2025 Jun 2;20(6):e0325177. doi: 10.1371/journal.pone.0325177. eCollection 2025.

ABSTRACT

Recent developments in artificial intelligence (AI) have introduced new technologies that can aid in detecting cognitive decline. This study developed a voice-based AI model that screens for cognitive decline using only a short conversational voice sample. The process involved collecting voice samples, applying machine learning (ML), and confirming accuracy through test data. The AI model extracts multiple voice features from the collected voice data to detect potential signs of cognitive impairment. Data labeling for ML was based on Mini-Mental State Examination scores: scores of 23 or lower were labeled as "cognitively declined (CD)," while scores above 24 were labeled as "cognitively normal (CN)." A fully coupled neural network architecture was employed for deep learning, using voice samples from 263 patients. Twenty voice samples, each comprising a one-minute conversation, were used for accuracy evaluation. The developed AI model achieved an accuracy of 0.950 in discriminating between CD and CN individuals, with a sensitivity of 0.875, specificity of 1.000, and an average area under the curve of 0.990. This voice AI model shows promise as a cognitive screening tool accessible via mobile devices, requiring no specialized environments or equipment, and can help detect CD, offering individuals the opportunity to seek medical attention.

PMID:40455724 | DOI:10.1371/journal.pone.0325177

PLoS One. 2025 Jun 2;20(6):e0325261. doi: 10.1371/journal.pone.0325261. eCollection 2025.

ABSTRACT

With the advancement of sensor technologies such as LiDAR and depth cameras, the significance of three-dimensional point cloud data in autonomous driving and environment sensing continues to increase.Point cloud registration stands as a fundamental task in constructing high-precision environmental models, with particular significance in overlapping regions where the accuracy of feature extraction and matching directly impacts registration quality. Despite advancements in deep learning approaches, existing methods continue to demonstrate limitations in extracting comprehensive features within these overlapping areas. This study introduces an innovative point cloud registration framework that synergistically combines the K-nearest neighbor (KNN) algorithm with a channel attention mechanism (CAM) to significantly enhance feature extraction and matching capabilities in overlapping regions. Additionally, by designing an effectiveness scoring network, the proposed method improves registration accuracy and enhances system robustness in complex scenarios. Comprehensive evaluations on the ModelNet40 dataset reveal that our approach achieves markedly superior performance metrics, demonstrating significantly lower root mean square error (RMSE) and mean absolute error (MAE) compared to established methods including iterative closest point (ICP), Robust & Efficient Point Cloud Registration using PointNet (PointNetLK), Go-ICP, fast global registration (FGR), deep closest point (DCP), self-supervised learning for a partial-to-partial registration (PRNet), and Iterative Distance-Aware Similarity Matrix Convolution (IDAM). This performance advantage is consistently maintained across various challenging conditions, including unseen shapes, novel categories, and noisy environments. Furthermore, additional experiments on the Stanford dataset validate the applicability and robustness of the proposed method for high-precision 3D shape registration tasks.

PMID:40455723 | DOI:10.1371/journal.pone.0325261