Deep learning
Localization and Classification of Adrenal Masses in Multiphase Computed Tomography: Retrospective Study
J Med Internet Res. 2025 Apr 24;27:e65937. doi: 10.2196/65937.
ABSTRACT
BACKGROUND: The incidence of adrenal incidentalomas is increasing annually, and most types of adrenal masses require surgical intervention. Accurate classification of common adrenal masses based on tumor computed tomography (CT) images by radiologists or clinicians requires extensive experience and is often challenging, which increases the workload of radiologists and leads to unnecessary adrenal surgeries. There is an urgent need for a fully automated, noninvasive, and precise approach for the identification and accurate classification of common adrenal masses.
OBJECTIVE: This study aims to enhance diagnostic efficiency and transform the current clinical practice of preoperative diagnosis of adrenal masses.
METHODS: This study is a retrospective analysis that includes patients with adrenal masses who underwent adrenalectomy from January 1, 2021, to May 31, 2023, at Center 1 (internal dataset), and from January 1, 2016, to May 31, 2023, at Center 2 (external dataset). The images include unenhanced, arterial, and venous phases, with 21,649 images used for the training set, 2406 images used for the validation set, and 12,857 images used for the external test set. We invited 3 experienced radiologists to precisely annotate the images, and these annotations served as references. We developed a deep learning-based adrenal mass detection model, Multi-Attention YOLO (MA-YOLO), which can automatically localize and classify 6 common types of adrenal masses. In order to scientifically evaluate the model performance, we used a variety of evaluation metrics, in addition, we compared the improvement in diagnostic efficacy of 6 doctors after incorporating model assistance.
RESULTS: A total of 516 patients were included. In the external test set, the MA-YOLO model achieved an intersection over union of 0.838, 0.885, and 0.890 for the localization of 6 types of adrenal masses in unenhanced, arterial, and venous phase CT images, respectively. The corresponding mean average precision for classification was 0.885, 0.913, and 0.915, respectively. Additionally, with the assistance of this model, the classification diagnostic performance of 6 radiologists and clinicians for adrenal masses improved. Except for adrenal cysts, at least 1 physician significantly improved diagnostic performance for the other 5 types of tumors. Notably, in the categories of adrenal adenoma (for senior clinician: P=.04, junior radiologist: P=.01, and senior radiologist: P=.01) and adrenal cortical carcinoma (junior clinician: P=.02, junior radiologist: P=.01, and intermediate radiologist: P=.001), half of the physicians showed significant improvements after using the model for assistance.
CONCLUSIONS: The MA-YOLO model demonstrates the ability to achieve efficient, accurate, and noninvasive preoperative localization and classification of common adrenal masses in CT examinations, showing promising potential for future applications.
PMID:40273442 | DOI:10.2196/65937
MIRACN: a residual convolutional neural network for predicting cell line specific functional regulatory variants
Brief Bioinform. 2025 Mar 4;26(2):bbaf196. doi: 10.1093/bib/bbaf196.
ABSTRACT
In post-genome-wide association study era, interpretation of noncoding variants remains a significant challenge due to their complexity and the limited understanding of their functions. Here, we developed MIRACN, a novel residual convolutional neural network designed to predict cell line-specific functional regulatory variants. By utilizing a substantial dataset from massively parallel reporter assays (MPRAs) and employing a multitask learning strategy, MIRACN was trained across seven distinct cell lines, attaining superior performance compared to existing methods, especially in predicting cell type specificity. Comparative evaluations on an independent MPRA test dataset demonstrated that MIRACN not only outperformed in identifying regulatory variants but also provided valuable insights into their cellular context-specific regulatory mechanisms. MIRACN is capable of not only providing scores for functional variants but also pinpointing the specific cell line in which these variants display their function. This enhancement has improved the resolution of current research on the functionality of noncoding variants and has paved the way for more precise diagnostic and therapeutic strategies.
PMID:40273430 | DOI:10.1093/bib/bbaf196
PathSynergy: a deep learning model for predicting drug synergy in liver cancer
Brief Bioinform. 2025 Mar 4;26(2):bbaf192. doi: 10.1093/bib/bbaf192.
ABSTRACT
Cancer is a major public health problem while liver cancer is the main cause of global cancer-related deaths. The previous study demonstrates that the 5-year survival rate for advanced liver cancer is only 30%. Few of the first-line targeted drugs including sorafenib and lenvatinib are available, which often develop resistance. Drug combination therapy is crucial for improving the efficacy of cancer therapy and overcoming resistance. However, traditional methods for discovering drug synergy are costly and time consuming. In this study, we developed a novel predicting model PathSynergy by integrating drug feature data, cell line data, drug-target interactions, and signaling pathways. PathSynergy combined the advantages of graph neural networks and pathway map mapping. Comparing with other baseline models, PathSynergy showed better performance in model classification, accuracy, and precision. Excitingly, six Food and Drug Administration (FDA)-approved drugs including pimecrolimus, topiramate, nandrolone_decanoate, fluticasone propionate, zanubrutinib, and levonorgestrel were predicted and validated to show synergistic effects with sorafenib or lenvatinib against liver cancer for the first time. In general, the PathSynergy model provides a new perspective to discover synergistic combinations of drugs and has broad application potential in the fields of drug discovery and personalized medicine.
PMID:40273429 | DOI:10.1093/bib/bbaf192
Shared-weight graph framework for comprehensive protein stability prediction across diverse mutation types
Brief Bioinform. 2025 Mar 4;26(2):bbaf190. doi: 10.1093/bib/bbaf190.
ABSTRACT
Research on protein stability changes is vital for understanding disease mechanisms and optimizing industrial enzymes. Protein thermal stability can be modified by variants leading to changes in ΔΔG values between wild-type and mutant proteins. Despite advances, most models focus on single-point mutations, overlooking multipoint and indel mutations. Typically, the single-point mutation is expected to have a relatively limited impact on the function of a protein, necessitating more drastic modifications to meet new challenges. Current methods for multipoint mutations yield poor results, and no method exists for any length of indel mutations. To address this, we introduce UniMutStab, a shared-graph convolutional network leveraging protein language models and residue interaction networks to access any type of mutation. An embedded edge weight module enhances the integration of residue node features and interactions, improving prediction accuracy. Trained on the "Mega-scale" dataset with ~780 000 mutations, UniMutStab surpasses existing methods in predicting protein stability changes. It is a purely sequence-based approach to predict arbitrary mutation types, demonstrating robust generalization across multiple tasks and potentially contributing significantly to protein engineering, personalized therapeutics, and diagnostic methodologies.
PMID:40273428 | DOI:10.1093/bib/bbaf190
DEKP: a deep learning model for enzyme kinetic parameter prediction based on pretrained models and graph neural networks
Brief Bioinform. 2025 Mar 4;26(2):bbaf187. doi: 10.1093/bib/bbaf187.
ABSTRACT
The prediction of enzyme kinetic parameters is crucial for screening enzymes with high catalytic efficiency and desired characteristics to catalyze natural or non-natural reactions. Data-driven machine learning models have been explored to reduce experimental cost and speed up the enzyme design process. However, the prediction performance is still subject to significant limitations due to the variance in sequence similarity between training and testing datasets. In this work, we introduce DEKP, an integrated deep learning approach enzyme kinetic parameter prediction. It leverages pretrained models of protein sequences and incorporates enhanced graph neural networks that provide comprehensive representation of protein structural features. This novel approach can effectively alleviate the performance degradation caused by sequence similarity variation. Moreover, it provides sensitive detection of changes in catalytic efficiency due to enzyme mutations. Experiments validate that DEKP outperforms existing models in predicting enzyme kinetic parameters. This work is expected to significantly improve the performance of the enzyme screening process and provide a robust tool for enzyme-directed evolution research.
PMID:40273427 | DOI:10.1093/bib/bbaf187
Intelligent recognition of human activities using deep learning techniques
PLoS One. 2025 Apr 24;20(4):e0321754. doi: 10.1371/journal.pone.0321754. eCollection 2025.
ABSTRACT
Recognition of Human Actions (HAR) Portrays a crucial significance in various applications due to its ability for analyzing behaviour of humans within videos. This research investigates HAR in Red, Green, and Blue, or RGB videos using frameworks for deep learning. The model's ensemble method integrates the forecasts from two models, 3D-AlexNet-RF and InceptionV3 Google-Net, to improve accuracy in recognizing human activities. Each model independently predicts the activity, and the ensembles method merges these predictions, often using voting or averaging, to produce a more accurate and reliable final classification. This approach leverages the advantages of each design, leading to enhanced performance recognition for activities. The performance of our ensemble framework is evaluated on our contesting HMDB51 dataset, known for its diverse human actions. Training the Inflated-3D (I3D) video classifiers using HMDB51 dataset, our system aims to improve patient care, enhance security, surveillances, Interaction between Humans and Computers, or HCI, and advance human-robot interaction. The ensemble model achieves exceptional results in every class, with an astounding aggregate accuracy of 99.54% accuracy, 97.94% precision, 97.94% recall, 99.56% specificity, 97.88% F1-Score, 95.43% IoU,97. 36% MCC and Cohen's Kappa 97.17%. These findings suggest that the ensemble model is highly effective & a powerful tool for HAR tasks. Multi-tiered ensembles boost wearable recognition, setting a new gold standard for healthcare, surveillance, and robotics.
PMID:40273193 | DOI:10.1371/journal.pone.0321754
Definer: A computational method for accurate identification of RNA pseudouridine sites based on deep learning
PLoS One. 2025 Apr 24;20(4):e0320077. doi: 10.1371/journal.pone.0320077. eCollection 2025.
ABSTRACT
Pseudouridine is an important modification site, which is widely present in a variety of non-coding RNAs and is involved in a variety of important biological processes. Studies have shown that pseudouridine is important in many biological functions such as gene expression, RNA structural stability, and various diseases. Therefore, accurate identification of pseudouridine sites can effectively explain the functional mechanism of this modification site. Due to the rapid increase of genomics data, traditional biological experimental methods to identify RNA modification sites can no longer meet the practical needs, and it is necessary to accurately identify pseudouridine sites from high-throughput RNA sequence data by computational methods. In this study, we propose a deep learning-based computational method, Definer, to accurately identify RNA pseudouridine loci in three species, Homo sapiens, Saccharomyces cerevisiae and Mus musculus. The method incorporates two sequence coding schemes, including NCP and One-hot, and then feeds the extracted RNA sequence features into a deep learning model constructed from CNN, GRU and Attention. The benchmark dataset contains data from three species, H. sapiens, S. cerevisiae and M. musculus, and the results using 10-fold cross-validation show that Definer significantly outperforms other existing methods. Meanwhile, the data sets of two species, H. sapiens and S. cerevisiae, were tested independently to further demonstrate the predictive ability of the model. In summary, our method, Definer, can accurately identify pseudouridine modification sites in RNA.
PMID:40273178 | DOI:10.1371/journal.pone.0320077
N-Beats architecture for explainable forecasting of multi-dimensional poultry data
PLoS One. 2025 Apr 24;20(4):e0320979. doi: 10.1371/journal.pone.0320979. eCollection 2025.
ABSTRACT
The agricultural economy heavily relies on poultry production, making accurate forecasting of poultry data crucial for optimizing revenue, streamlining resource utilization, and maximizing productivity. This research introduces a novel application of the N-BEATS architecture for multi-dimensional poultry data forecasting with enhanced interpretability through an integrated Explainable AI (XAI) framework. Leveraging its advanced capabilities in time series modeling, N-BEATS is applied to predict multiple facets of poultry disease diagnostics using a multivariate dataset comprising key environmental parameters. The methodology empowers decision-making in poultry farm management by providing transparent and interpretable forecasts. Experimental results demonstrate that N-BEATS outperforms conventional deep learning models, including LSTM, GRU, RNN, and CNN, across various error metrics, achieving MAE of 0.172, RMSE of 0.313, MSLE of 0.042, R-squared of 0.034, and RMSLE of 0.204. The positive R-squared value indicates the model's robustness against underfitting and overfitting, surpassing the performance of other models with negative R-squared values. This study establishes N-BEATS as a superior and interpretable solution for complex, multi-dimensional forecasting challenges in poultry production, with significant implications for enhancing predictive analytics in agriculture.
PMID:40273069 | DOI:10.1371/journal.pone.0320979
Leveraging deep learning models to increase the representation of nomadic pastoralists in health campaigns and demographic surveillance
PLOS Glob Public Health. 2025 Apr 24;5(4):e0004018. doi: 10.1371/journal.pgph.0004018. eCollection 2025.
ABSTRACT
Nomadic pastoralists are systematically underrepresented in the planning of health services and frequently missed by health campaigns due to their mobility. Previous studies have developed novel geospatial methods to address these challenges but rely on manual techniques that are too time and resource-intensive to scale on a national or regional level. To address this gap, we developed a computer vision-based approach to automatically locate active nomadic pastoralist settlements from satellite imagery. We curated labeled datasets of satellite images capturing approximately 1,000 historically active settlements in the Omo Valley of Ethiopia and the Samburu County of Kenya to train and evaluate deep learning models, studying their robustness to low spatial resolutions and limits in labeled training data. Using a novel training strategy that leveraged public road and water infrastructure data, we closed performance gaps introduced by shortages in labeled settlement data. We deployed our best model on a region spanning 5,400 square kilometers in the Omo Valley of Ethiopia, resulting in the identification of historical settlements with a 270-fold reduction in manual review volume. Our work serves as a promising framework for automating the localization of nomadic pastoralist settlements at a national scale for health campaigns and demographic surveillance.
PMID:40273062 | DOI:10.1371/journal.pgph.0004018
Elevator fault precursor prediction based on improved LSTM-AE algorithm and TSO-VMD denoising technique
PLoS One. 2025 Apr 24;20(4):e0320566. doi: 10.1371/journal.pone.0320566. eCollection 2025.
ABSTRACT
This study proposes an advanced elevator fault precursor prediction method integrating Variational Mode Decomposition (VMD), Bidirectional Long Short-Term Memory (BILSTM), and an Autoencoder with an Attention Mechanism (AEAM), collectively referred to as the VMD-BILSTM-AEAM algorithm. This model addresses the challenges of feature redundancy and noise interference in elevator operation data, improving the stability and accuracy of fault predictions. Using a dataset of elevator operation parameters, including current, voltage, and running speed, the model utilizes the Attribute Correlation Density Ranking (ACDR) method for feature selection and the TSO-optimized VMD for denoising, enhancing data quality. Cross-validation and statistical analyses, including confidence interval calculations, were employed to validate the robustness of the model. The results demonstrate that the VMD-BILSTM-AEAM algorithm achieves a mean True Positive Rate (TPR) of 0.919 with a 95% confidence interval of 0.915 to 0.924, a mean False Positive Rate (FPR) of 0.090 with a 95% confidence interval of 0.087 to 0.092, and a mean Area Under the Curve (AUC) of 0.919 with a 95% confidence interval of 0.915 to 0.923. These performance metrics indicate a significant improvement over traditional and other deep learning models, confirming the model's superiority in predictive maintenance of elevators. The robust capability of the VMD-BILSTM-AEAM algorithm to accurately process and analyze time-series data, even in the presence of noise, highlights its potential for broader applications in predictive maintenance and fault detection across various domains.
PMID:40273057 | DOI:10.1371/journal.pone.0320566
Visceral Fat Quantified by a Fully Automated Deep-Learning Algorithm and Risk of Incident and Recurrent Diverticulitis
Dis Colon Rectum. 2025 Mar 4. doi: 10.1097/DCR.0000000000003677. Online ahead of print.
ABSTRACT
BACKGROUND: Obesity is a risk factor for diverticulitis. However, it remains unclear whether visceral fat area, a more precise measurement of abdominal fat, is associated with the risk of diverticulitis.
OBJECTIVE: To estimate the risk of incident and recurrent diverticulitis according to visceral fat area.
DESIGN: A retrospective cohort study.
SETTINGS: The Mass General Brigham Biobank.
PATIENTS: 6,654 patients who underwent abdominal CT for clinical indications and had no diagnosis of diverticulitis, inflammatory bowel disease, or cancer before the scan.
MAIN OUTCOME MEASURES: Visceral fat area, subcutaneous fat area, and skeletal muscle area were quantified using a deep-learning model applied to abdominal CT. The main exposures were z-scores of body composition metrics, normalized by age, sex, and race. Diverticulitis cases were defined with the ICD codes for the primary or admitting diagnosis from the electronic health records. The risks of incident diverticulitis, complicated diverticulitis, and recurrent diverticulitis requiring hospitalization according to quartiles of body composition metrics z-scores were estimated.
RESULTS: A higher visceral fat area z-score was associated with an increased risk of incident diverticulitis (multivariable HR comparing the highest versus lowest quartile, 2.09; 95% CI, 1.48-2.95; P for trend <.0001), complicated diverticulitis (HR, 2.56; 95% CI, 1.10-5.99; P for trend = .02), and recurrence requiring hospitalization (HR, 2.76; 95% CI, 1.15-6.62; P for trend = .03). The association between visceral fat area and diverticulitis was not materially different among different strata of body mass index. Subcutaneous fat area and skeletal muscle area were not significantly associated with diverticulitis.
LIMITATIONS: The study population was limited to individuals who underwent CT scans for medical indication.
CONCLUSION: Higher visceral fat area derived from CT was associated with incident and recurrent diverticulitis. Our findings provide insight into the underlying pathophysiology of diverticulitis and may have implications for preventive strategies. See Video Abstract.
PMID:40272787 | DOI:10.1097/DCR.0000000000003677
perfDSA: Automatic Perfusion Imaging in Cerebral Digital Subtraction Angiography
Int J Comput Assist Radiol Surg. 2025 Apr 24. doi: 10.1007/s11548-025-03359-4. Online ahead of print.
ABSTRACT
PURPOSE: Cerebral digital subtraction angiography (DSA) is a standard imaging technique in image-guided interventions for visualizing cerebral blood flow and therapeutic guidance thanks to its high spatio-temporal resolution. To date, cerebral perfusion characteristics in DSA are primarily assessed visually by interventionists, which is time-consuming, error-prone, and subjective. To facilitate fast and reproducible assessment of cerebral perfusion, this work aims to develop and validate a fully automatic and quantitative framework for perfusion DSA.
METHODS: We put forward a framework, perfDSA, that automatically generates deconvolution-based perfusion parametric images from cerebral DSA. It automatically extracts the arterial input function from the supraclinoid internal carotid artery (ICA) and computes deconvolution-based perfusion parametric images including cerebral blood volume (CBV), cerebral blood flow (CBF), mean transit time (MTT), and Tmax.
RESULTS: On a DSA dataset with 1006 patients from the multicenter MR CLEAN registry, the proposed perfDSA achieves a Dice of 0.73(±0.21) in segmenting the supraclinoid ICA, resulting in high accuracy of arterial input function (AIF) curves similar to manual extraction. Moreover, some extracted perfusion images show statistically significant associations (P=2.62e - 5) with favorable functional outcomes in stroke patients.
CONCLUSION: The proposed perfDSA framework promises to aid therapeutic decision-making in cerebrovascular interventions and facilitate discoveries of novel quantitative biomarkers in clinical practice. The code is available at https://github.com/RuishengSu/perfDSA .
PMID:40272658 | DOI:10.1007/s11548-025-03359-4
Role of artificial intelligence in advancing immunology
Immunol Res. 2025 Apr 24;73(1):76. doi: 10.1007/s12026-025-09632-7.
ABSTRACT
Artificial intelligence (AI) has revolutionized various biomedical fields, particularly immunology, by enhancing vaccine development, immunotherapies, and allergy treatments. AI helps identify potential vaccine candidates and predict how the body reacts to different antigens based on a vast number of genomic sequences and protein structures. AI can help cancer patients by analyzing their data and offering personalized immunotherapies. AI has also advanced the field of allergy by identifying potential allergens and predicting allergic reactions based on patient genetic and environmental factors. AI could also help diagnose multiple immunological diseases, including autoimmune diseases and immunodeficiencies, by analyzing patient history and laboratory results. AI has deepened our understanding of the human genome by providing numerous amounts of data from DNA sequences previously believed to be nonfunctional. Through machine learning and deep learning, many laborious research tasks, such as screening for DNA mutations, can be efficiently performed in a short amount of time. AI and machine learning are significantly advancing biomedical science in significant areas, including research and industry. This review discusses the latest AI-based tools that can be utilized in the field of immunology. AI tools significantly advance the field of medical research and healthcare by enabling new scientific discoveries and facilitating rapid clinical diagnosis.
PMID:40272607 | DOI:10.1007/s12026-025-09632-7
Relationships Between Retinal Vascular Characteristics and Systemic Indicators in Patients With Diabetes Mellitus
Invest Ophthalmol Vis Sci. 2025 Apr 1;66(4):72. doi: 10.1167/iovs.66.4.72.
ABSTRACT
PURPOSE: To develop a deep learning method for vessel segmentation in fundus images, measure retinal vessels, and study the connection between retinal vascular features and systemic indicators in diabetic patients.
METHODS: We conducted a study on patients with diabetes mellitus (DM) at various stages of diabetic retinopathy (DR) using data from the Joint Asia Diabetes Evaluation (JADE) Register. All participants underwent comprehensive clinical assessments, including anthropometric measurements, laboratory tests, and fundus photography, during each follow-up visit (2.81 average follow-up visits). A custom U-Net deep learning model utilizing a variety of open-source datasets was developed for the segmentation and measurement of retinal vessels. We investigated the relationship between systemic indicators and the severity of DR, analyzing the correlation coefficients between systemic indicators and retinal vascular characteristics.
RESULTS: We enrolled a total of 637 patients diagnosed with DM and collected 3575 series of photographs for analysis. Some of the systemic indicators and retinal vascular metrics, including central retinal arteriolar equivalent, central retinal venular equivalent, arteriole-to-venule ratio, and fractal dimension, were significantly correlated with the severity of diabetic retinopathy (P < 0.05). Some physical characteristics, hematological parameters, renal function parameters, metabolism-related parameters, biochemical markers such as folic acid and fasting insulin, liver enzymes, and macrovascular indicators were significantly correlated with certain retinal vascular metrics (P < 0.05).
CONCLUSIONS: Multiple systemic indicators were identified as significantly associated with the advancement of diabetic retinopathy and retinal vascular metrics. Utilizing deep learning techniques for vessel segmentation and measurement on color fundus photographs can help elucidate the connections between retinal vascular characteristics and systemic indicators.
PMID:40272369 | DOI:10.1167/iovs.66.4.72
Artificial Intelligence in Panoramic Radiography Interpretation: A Glimpse into the State-of-the-Art Radiologic Examination Method
Int J Comput Dent. 2025 Apr 24;0(0):0. doi: 10.3290/j.ijcd.b6173229. Online ahead of print.
ABSTRACT
AIM: Panoramic radiography is a frequently utilized imaging technique in standard dental examinations and provides many advantages. In this context, studies have been conducted to develop tools to assist physicians in clinical practice by using deep learning models to interpret panoramic radiography images. However, studies in the existing literature have generally addressed these conditions separately and studies that develop a multiclass diagnostic charting model that can detect and segment all these conditions are very limited. Therefore, the aim of this study to develop a deep learning model that can accurately evaluate and segment various dental issues and anatomical structures in panoramic radiographs obtained from different radiography devices and settings.
MATERIALS AND METHODS: Panoramic radiographs were labelled for 33 different conditions in the categories of dental problems, dental restorations, dental implants, anatomical landmarks, periodontal conditions, jaw pathologies and periapical lesions. A YOLO-v8 model was employed to develop an artificial intelligence model for each labelling. A confusion matrix was utilised to successfully evaluate the developed models.
RESULTS: The algorithm achieved a precision value of 0.99-1 in accurately detecting various dental features, such as adult tooth numbering, filling, dental implants, dental pulp, root canal filling, mandibular canal, mandibular condyle, mandible, and pharyngeal airway. With respect to sensitivity, the adult tooth numbering, dental implants, mandibular canal, maxillary sinus, mandibular condyle, angulus mandible, nasal septum, mandible, and hard palate showed the highest values of 0.99-1. The F1-score reached the highest value of 0.99-1 for the root canal filling, adult tooth numbering, dental implants, mandibular canal, mandibular condyle, angulus mandible, mandible, and pharyngeal airway.
CONCLUSION: Artificial intelligence based on convolutional neural networks has a remarkable ability to detect different conditions observed in regular clinical evaluations in panoramic radiographs, displaying excellent performance. Based on these findings, it can be confidently stated that deep learning-based models has great potential to improve routine clinical practices for physicians.
PMID:40272192 | DOI:10.3290/j.ijcd.b6173229
Intelligent Diagnosis of Cervical Lymph Node Metastasis Using a CNN Model
J Dent Res. 2025 Apr 24:220345251322508. doi: 10.1177/00220345251322508. Online ahead of print.
ABSTRACT
Lymph node (LN) metastasis is a prevalent cause of recurrence in oral squamous cell carcinoma (OSCC). However, accurately identifying metastatic LNs (LNs+) remains challenging. This prospective clinical study aims to test the effectiveness of our convolutional neural network (CNN) model for identifying OSCC cervical LN+ in contrast-enhanced computed tomography (CECT) in clinical practice. A CNN model was developed and trained using a dataset of 8,380 CECT images from previous OSCC patients. It was then prospectively validated on 17,777 preoperative CECT images from 354 OSCC patients between October 17, 2023, and August 31, 2024. The model's predicted LN results were provided to the surgical team without influencing surgical or treatment plans. During surgery, the predicted LN+ were identified and sent for separate pathological examination. The accuracy of the model's predictions was compared with those of human experts and verified against pathology reports. The capacity of the model to assist radiologists in LN+ diagnosis was also assessed. The CNN model was trained over 40 epochs and successfully validated after each. Compared with human experts (2 radiologists, 2 surgeons, and 2 students), the CNN model achieved higher sensitivity (81.89% vs. 81.48%, 46.91%, 50.62%), specificity (99.31% vs. 99.15%, 98.36%, 96.27%), LN+ accuracy (76.19% vs. 75.43%, P = 0.854; 40.64%, P < 0.001; 37.44%, P < 0.001), and clinical accuracy (86.16% vs. 83%, 61%, 56%). With the model's assistance, the radiologists surpassed both the previous predictive results without the model's support and the model's performance alone. The CNN model demonstrated an accuracy comparable to that of radiologists in identifying, locating, and predicting cervical LN+ in OSCC patients. Furthermore, the model has the potential to assist radiologists in making more accurate diagnoses.
PMID:40271993 | DOI:10.1177/00220345251322508
Deep Learning-Assisted Design for High-Q-Value Dielectric Metasurface Structures
Materials (Basel). 2025 Mar 29;18(7):1554. doi: 10.3390/ma18071554.
ABSTRACT
Optical sensing technologies play a crucial role in various fields such as biology, medicine, and food safety by measuring changes in material properties, such as the refractive index, light absorption, and scattering. Dielectric metasurfaces, with their subwavelength-scale geometric features and the ability to achieve high-quality-factor (Q-value) resonances through specific meta-atom designs, offer a new avenue for achieving faster and more sensitive material detection. The resonant wavelength, as one of the key indicators in meta-atom design, is usually determined using traditional solving methods such as electromagnetic simulations, which, although capable of providing high-precision prediction results, suffer from slow computational speed and long processing times. To address this issue, this paper proposes a forward prediction network for the amplitude spectrum of dielectric metasurfaces. Test results demonstrated that the mean square error of this network was consistently less than 10-3, and the neural network required less than 1 s, indicating its high-precision prediction capability. Furthermore, we employed transfer learning to apply this network to predict the near-infrared transmission spectra of high-Q-value resonant dielectric metasurfaces, achieving significant effectiveness. This method greatly enhanced the efficiency of metasurface design, and the designed network could serve as a universal backbone model for the forward prediction of spectral responses for other types of dielectric metasurfaces.
PMID:40271794 | DOI:10.3390/ma18071554
Advanced Thermal Imaging Processing and Deep Learning Integration for Enhanced Defect Detection in Carbon Fiber-Reinforced Polymer Laminates
Materials (Basel). 2025 Mar 25;18(7):1448. doi: 10.3390/ma18071448.
ABSTRACT
Carbon fiber-reinforced polymer (CFRP) laminates are widely used in aerospace, automotive, and infrastructure industries due to their high strength-to-weight ratio. However, defect detection in CFRP remains challenging, particularly in low signal-to-noise ratio (SNR) conditions. Conventional segmentation methods often struggle with noise interference and signal variations, leading to reduced detection accuracy. In this study, we evaluate the impact of thermal image preprocessing on improving defect segmentation in CFRP laminates inspected via pulsed thermography. Polynomial approximations and first- and second-order derivatives were applied to refine thermographic signals, enhancing defect visibility and SNR. The U-Net architecture was used to assess segmentation performance on datasets with and without preprocessing. The results demonstrated that preprocessing significantly improved defect detection, achieving an Intersection over Union (IoU) of 95% and an F1-Score of 99%, outperforming approaches without preprocessing. These findings emphasize the importance of preprocessing in enhancing segmentation accuracy and reliability, highlighting its potential for advancing non-destructive testing techniques across various industries.
PMID:40271635 | DOI:10.3390/ma18071448
Mixed Outcomes in Recombination Rates After Domestication: Revisiting Theory and Data
Mol Ecol. 2025 Apr 24:e17773. doi: 10.1111/mec.17773. Online ahead of print.
ABSTRACT
The process of domestication has altered many phenotypes. Selection on these phenotypes has long been hypothesised to indirectly select for increases in the genome-wide recombination rate. This hypothesis is potentially consistent with theory on the evolution of the recombination rate, but empirical support has been unclear. We review relevant theory, lab-based experiments, and data comparing recombination rates in wild progenitors and their domesticated counterparts. We utilise population sequencing data and a deep learning method to infer genome-wide recombination rates for new comparisons of chicken/red junglefowl, sheep/mouflon, and goat/bezoar. We find evidence of increased recombination in domesticated goats compared to bezoars but more mixed results in chicken and generally decreased recombination in domesticated sheep compared to mouflon. Our results add to a growing body of literature in plants and animals that finds no consistent evidence of an increase in genome-wide recombination with domestication.
PMID:40271548 | DOI:10.1111/mec.17773
A recognition model for winter peach fruits based on improved ResNet and multi-scale feature fusion
Front Plant Sci. 2025 Apr 9;16:1545216. doi: 10.3389/fpls.2025.1545216. eCollection 2025.
ABSTRACT
With the continuous advancement of modern agricultural technologies, the demand for precision fruit-picking techniques has been increasing. This study addresses the challenge of accurate recognition and harvesting of winter peaches by proposing a novel recognition model based on the residual network (ResNet) architecture-WinterPeachNet-aimed at enhancing the accuracy and efficiency of winter peach detection, even in resource-constrained environments. The WinterPeachNet model achieves a comprehensive improvement in network performance by integrating depthwise separable inverted bottleneck ResNet (DIBResNet), bidirectional feature pyramid network (BiFPN) structure, GhostConv module, and the YOLOv11 detection head (v11detect). The DIBResNet module, based on the ResNet architecture, introduces an inverted bottleneck structure and depthwise separable convolution technology, enhancing the depth and quality of feature extraction while effectively reducing the model's computational complexity. The GhostConv module further improves detection accuracy by reducing the number of convolution kernels. Additionally, the BiFPN structure strengthens the model's ability to detect objects of different sizes by fusing multi-scale feature information. The introduction of v11detect further optimizes object localization accuracy. The results show that the WinterPeachNet model achieves excellent performance in the winter peach detection task, with P = 0.996, R = 0.996, mAP50 = 0.995, and mAP50-95 = 0.964, demonstrating the model's efficiency and accuracy in the winter peach detection task. The high efficiency of the WinterPeachNet model makes it highly adaptable in resource-constrained environments, enabling effective object detection at a relatively low computational cost.
PMID:40271441 | PMC:PMC12014684 | DOI:10.3389/fpls.2025.1545216