Sci Rep. 2025 Feb 21;15(1):6360. doi: 10.1038/s41598-025-85582-2.

ABSTRACT

As a high-risk production unit, natural gas supply enterprises are increasingly recognizing the need to enhance production safety management. Traditional process warning methods, which rely on fixed alarm values, often fail to adequately account for dynamic changes in the production process. To address this issue, this study utilizes deep learning techniques to enhance the accuracy and reliability of natural gas load forecasting. By considering the benefits and feasibility of integrating multiple models, a VMD-CNN-LSTM-Self-Attention interval prediction method was innovatively proposed and developed. Empirical research was conducted using data from natural gas field station outgoing loads. The primary model constructed is a deep learning model for interval prediction of natural gas loads, which implements a graded alarm mechanism based on 85%, 90%, and 95% confidence intervals of real-time observations. This approach represents a novel strategy for enhancing enterprise safety production management. Experimental results demonstrate that the proposed method outperforms traditional warning models, reducing MAE, MAPE, MESE, and REMS by 1.13096 m3/h, 1.3504%, 7.6363 m3/h, 1.6743 m3/h, respectively, while improving R2 by 0.04698. These findings are expected to offer valuable insights for enhancing safe production management in the natural gas industry and provide new perspectives for the industry's digital and intelligent transformation.

PMID:39984509 | DOI:10.1038/s41598-025-85582-2

Nat Commun. 2025 Feb 21;16(1):1838. doi: 10.1038/s41467-025-57072-6.

ABSTRACT

Inferring cell spatial profiles from histology images is critical for cancer diagnosis and treatment in clinical settings. In this study, we report a weakly-supervised deep-learning method, HistoCell, to directly infer super-resolution cell spatial profiles consisting of cell types, cell states and their spatial network from histology images at the single-nucleus-level. Benchmark analysis demonstrates that HistoCell robustly achieves state-of-the-art performance in terms of cell type/states prediction solely from histology images across multiple cancer tissues. HistoCell can significantly enhance the deconvolution accuracy for the spatial transcriptomics data and enable accurate annotation of subtle cancer tissue architectures. Moreover, HistoCell is applied to de novo discovery of clinically relevant spatial organization indicators, including prognosis and drug response biomarkers, across diverse cancer types. HistoCell also enable image-based screening of cell populations that drives phenotype of interest, and is applied to discover the cell population and corresponding spatial organization indicators associated with gastric malignant transformation risk. Overall, HistoCell emerges as a powerful and versatile tool for cancer studies in histology image-only cohorts.

PMID:39984438 | DOI:10.1038/s41467-025-57072-6

BMC Med. 2025 Feb 21;23(1):97. doi: 10.1186/s12916-025-03887-5.

ABSTRACT

BACKGROUND: Early detection of cancer can help patients with more effective treatments and result in better prognosis. Unfortunately, established cancer screening technologies are limited for use, especially for multi-cancer early detection. In this study, we described a serum-based platform integrating surface-enhanced Raman spectroscopy (SERS) technology with resampling strategy, feature dimensionality enhancement, deep learning and interpretability analysis methods for sensitive and accurate pan-cancer screening.

METHODS: Totally, 1655 early-stage patients with breast cancer (BC, n = 569), lung cancer (LC, n = 513), thyroid cancer (TC, n = 220), colorectal cancer (CC, n = 215), gastric cancer (GC, n = 100), esophageal cancer (EC, n = 38), and 1896 healthy controls (HC) were enrolled. The serum SERS spectra were obtained from each participant. Data dimension enhancement was conducted by heatmap transformation and continuous wavelet transform (CWT). The dimensionalization SERS spectral data were subsequently analyzed by residual neural network (ResNet) as convolutional neural network (CNN) algorithm. Class activation mapping (CAM) method was performed to elucidate the potential biological significance of spectral data classification.

RESULTS: All participants were divided into a training set and a test set with a ratio of 7:3. The BorderlineSMOTE method was selected as the most appropriate resampling strategy and the deep neural network (DNN) model achieved desirable performance among all groups (accuracy rate: 93.15%, precision rate: 88:46%, recall rate: 85.68%, and F1-score: 86.98%), with the generated AUC values of 0.991 for HC, 0.995 for BC, 0.979 for LC, 0.996 for TC, 0.994 for CC, 0.982 for GC, and 0.941 for EC, respectively. Furthermore, the combination use of SERS spectra data and ResNet (form of heatmap) were also capable of effectively distinguishing different categories and making accurate predictions (accuracy rate: 94.75%, precision rate: 89.02, recall rate: 86.97, and F1-score: 87.88), with the AUC values of 0.996 for HC, 0.995 for BC, 0.988 for LC, 0.999 for TC, 0.993 for CC, 0.985 for GC, and 0.940 for EC, respectively. Additionally, strong wave number range of the spectral data was observed in the CAM analysis.

CONCLUSIONS: Our study has offered a highly effective serum SERS-based approach for multi-cancer early detection, which might shed new light on cancer screening in clinical practice.

PMID:39984977 | DOI:10.1186/s12916-025-03887-5

BMC Ophthalmol. 2025 Feb 21;25(1):86. doi: 10.1186/s12886-025-03908-0.

ABSTRACT

BACKGROUND: Morphological changes in the retina are crucial and serve as valuable references in the clinical diagnosis of ophthalmic and cardiovascular diseases. However, the retinal vascular structure is complex, making manual segmentation time-consuming and labor-intensive.

METHODS: This paper proposes a retinal segmentation network that integrates feature channel attention and the Convolutional Block Attention Module (CBAM) attention within the U2Net model. First, a feature channel attention module is introduced into the RSU (Residual Spatial Unit) block of U2Net, forming an Attention-RSU block, which focuses more on significant areas during feature extraction and suppresses the influence of noise; Second, a Spatial Attention Module (SAM) is introduced into the high-resolution module of Attention-RSU to enrich feature extraction from both spatial and channel dimensions, and a Channel Attention Module (CAM) is integrated into the lowresolution module of Attention-RSU, which uses dual channel attention to reduce detail loss.Finally, dilated convolution is applied during the upscaling and downscaling processes to expand the receptive field in low-resolution states, allowing the model to better integrate contextual information.

RESULTS: The evaluation across multiple clinical datasets demonstrated excellent performance on various metrics, with an accuracy (ACC) of 98.71%.

CONCLUSION: The proposed Network is general enough and we believe it can be easily extended to other medical image segmentation tasks where large scale variation and complicated features are the main challenges.

PMID:39984892 | DOI:10.1186/s12886-025-03908-0

BMC Biol. 2025 Feb 21;23(1):47. doi: 10.1186/s12915-025-02148-4.

ABSTRACT

BACKGROUND: Long non-coding RNA (LncRNA) play pivotal roles in various cellular processes, and elucidating their subcellular localization can offer crucial insights into their functional significance. Accurate prediction of lncRNA subcellular localization is of paramount importance. Despite numerous computational methods developed for this purpose, existing approaches still encounter challenges stemming from the complexity of data representation and the difficulty in capturing nucleotide distribution information within sequences.

RESULTS: In this study, we propose a novel deep learning-based model, termed MGBLncLoc, which incorporates a unique multi-class encoding technique known as generalized encoding based on the Distribution Density of Multi-Class Nucleotide Groups (MCD-ND). This encoding approach enables more precise reflection of nucleotide distributions, distinguishing between constant and discriminative regions within sequences, thereby enhancing prediction performance. Additionally, our deep learning model integrates advanced neural network modules, including Multi-Dconv Head Transposed Attention, Gated-Dconv Feed-forward Network, Convolutional Neural Network, and Bidirectional Gated Recurrent Unit, to comprehensively exploit sequence features of lncRNA.

CONCLUSIONS: Comparative analysis against commonly used sequence feature encoding methods and existing prediction models validates the effectiveness of MGBLncLoc, demonstrating superior performance. This research offers novel insights and effective solutions for predicting lncRNA subcellular localization, thereby providing valuable support for related biological investigations.

PMID:39984880 | DOI:10.1186/s12915-025-02148-4

BMC Med Imaging. 2025 Feb 21;25(1):56. doi: 10.1186/s12880-024-01517-9.

ABSTRACT

PURPOSE: To develop MRI-based deep learning (DL) models for distinguishing sinonasal squamous cell carcinoma (SCC), adenoid cystic carcinoma (ACC) and olfactory neuroblastoma (ONB) and to evaluate whether the DL models could improve the diagnostic performance of Senior radiologist (SR) and Junior radiologist (JR).

METHODS: This retrospective analysis consisted of 465 patients (229 sinonasal SCCs, 128 ACCs and 108 ONBs). The training and validation cohorts included 325 and 47 patients and the independent external testing cohort consisted of 93 patients. MRI images included T2-weighted image (T2WI), contrast-enhanced T1-weighted image (CE-T1WI) and apparent diffusion coefficient (ADC). We analyzed the conventional MRI features to choose the independent predictors and built the conventional MRI model. Then we compared the macro- and micro- area under the curves (AUCs) of different sequences and different DL networks to formulate the best DL model [artificial intelligence (AI) model scheme]. With AI assistance, we observed the diagnostic performances between SR and JR. The diagnostic efficacies of SR and JR were assessed by accuracy, Recall, precision, F1-Score and confusion matrices.

RESULTS: The independent predictors of conventional MRI included intensity on T2WI and intracranial invasion of sinonasal malignancies. With ExtraTrees (ET) classier, the conventional MRI model owned AUC of 78.8%. For DL models, ResNet101 network showed better performance than ResNet50 and DensNet121, especially for the mean fusion sequence (macro-AUC = 0.892, micro-AUC = 0.875, Accuracy = 0.810), and also good for the ADC sequence (macro-AUC = 0.872, micro-AUC = 0.874, Accuracy = 0.814). Grad-CAM showed that DL models focused on solid component of lesions. With the best AI scheme (ResNet101-mean sequence-based DL model) assistance, the diagnosis performances of SR (accuracy = 0.957, average Recall = 0.962, precision = 0.955, F1-Score = 0.957) and JR (accuracy = 0.925, average Recall = 0.917, precision = 0.931, F1-Score = 0.923) were significantly improved.

CONCLUSION: The ResNet101 network with mean sequence based DL model could effectively differential between sinonasal SCC, ACC and ONB and improved the diagnostic performances of both senior and junior radiologists.

PMID:39984860 | DOI:10.1186/s12880-024-01517-9

BMC Med Res Methodol. 2025 Feb 21;25(1):45. doi: 10.1186/s12874-025-02463-y.

ABSTRACT

BACKGROUND: This scoping review systematically maps externally validated machine learning (ML)-based models in cancer patient care, quantifying their performance, and clinical utility, and examining relationships between models, cancer types, and clinical decisions. By synthesizing evidence, this study identifies, strengths, limitations, and areas requiring further research.

METHODS: The review followed the Joanna Briggs Institute's methodology, Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews guidelines, and the Population, Concept, and Context mnemonic. Searches were conducted across Embase, IEEE Xplore, PubMed, Scopus, and Web of Science (January 2014-September 2022), targeting English-language quantitative studies in Q1 journals (SciMago Journal and Country Ranking > 1) that used ML to evaluate clinical outcomes for human cancer patients with commonly available data. Eligible models required external validation, clinical utility assessment, and performance metric reporting. Studies involving genetics, synthetic patients, plants, or animals were excluded. Results were presented in tabular, graphical, and descriptive form.

RESULTS: From 4023 deduplicated abstracts and 636 full-text reviews, 56 studies (2018-2022) met the inclusion criteria, covering diverse cancer types and applications. Convolutional neural networks were most prevalent, demonstrating high performance, followed by gradient- and decision tree-based algorithms. Other algorithms, though underrepresented, showed promise. Lung and digestive system cancers were most frequently studied, focusing on diagnosis and outcome predictions. Most studies were retrospective and multi-institutional, primarily using image-based data, followed by text-based and hybrid approaches. Clinical utility assessments involved 499 clinicians and 12 tools, indicating improved clinician performance with AI assistance and superior performance to standard clinical systems.

DISCUSSION: Interest in ML-based clinical decision-making has grown in recent years alongside increased multi-institutional collaboration. However, small sample sizes likely impacted data quality and generalizability. Persistent challenges include limited international validation across ethnicities, inconsistent data sharing, disparities in validation metrics, and insufficient calibration reporting, hindering model comparison reliability.

CONCLUSION: Successful integration of ML in oncology decision-making requires standardized data and methodologies, larger sample sizes, greater transparency, and robust validation and clinical utility assessments.

OTHER: Financed by FCT-Fundação para a Ciência e a Tecnologia (Portugal, project LA/P/0063/2020, grant 2021.09040.BD) as part of CSS's Ph.D. This work was not registered.

PMID:39984835 | DOI:10.1186/s12874-025-02463-y

NPJ Digit Med. 2025 Feb 21;8(1):118. doi: 10.1038/s41746-025-01507-3.

ABSTRACT

This study forms the basis of a digital twin system of the knee joint, using advanced quantitative MRI (qMRI) and machine learning to advance precision health in osteoarthritis (OA) management and knee replacement (KR) prediction. We combined deep learning-based segmentation of knee joint structures with dimensionality reduction to create an embedded feature space of imaging biomarkers. Through cross-sectional cohort analysis and statistical modeling, we identified specific biomarkers, including variations in cartilage thickness and medial meniscus shape, that are significantly associated with OA incidence and KR outcomes. Integrating these findings into a comprehensive framework represents a considerable step toward personalized knee-joint digital twins, which could enhance therapeutic strategies and inform clinical decision-making in rheumatological care. This versatile and reliable infrastructure has the potential to be extended to broader clinical applications in precision health.

PMID:39984725 | DOI:10.1038/s41746-025-01507-3

Eur J Trauma Emerg Surg. 2025 Feb 21;51(1):119. doi: 10.1007/s00068-025-02793-y.

ABSTRACT

PURPOSE: The application of artificial intelligence (AI) in healthcare has seen widespread implementation, with numerous studies highlighting the development of robust algorithms. However, limited attention has been given to the secure utilization of raw data for medical model training, and its subsequent impact on clinical decision-making and real-world applications. This study aims to assess the feasibility and effectiveness of an advanced diagnostic model that integrates blockchain technology and AI for the identification of tibial plateau fractures (TPFs) in emergency settings.

METHOD: In this study, blockchain technology was utilized to construct a distributed database for trauma orthopedics, images collected from three independent hospitals for model training, testing, and internal validation. Then, a distributed network combining blockchain and deep learning was developed for the detection of TPFs, with model parameters aggregated across multiple nodes to enhance accuracy. The model's performance was comprehensively evaluated using metrics including accuracy, sensitivity, specificity, F1 score, and the area under the receiver operating characteristic curve (AUC). In addition, the performance of the centralized model, the distributed AI model, clinical orthopedic attending physicians, and AI-assisted attending physicians was tested on an external validation dataset.

RESULTS: In the testing set, the accuracy of our distributed model was 0.9603 [95% CI (0.9598, 0.9605)] and the AUC was 0.9911 [95% CI (0.9893, 0.9915)] for TPF detection. In the external validation set, the accuracy reached 0.9636 [95% CI (0.9388, 0.9762)], was slightly higher than that of the centralized YOLOv8n model at 0.9632 [95% CI (0.9387, 0.9755)] (p > 0.05), and exceeded the orthopedic physician at 0.9291 [95% CI (0.9002, 0.9482)] and radiology attending physician at 0.9175 [95% CI (0.8891, 0.9393)], with a statistically significant difference (p < 0.05). Additionally, the centralized model (4.99 ± 0.01 min) had shorter diagnosis times compared to the orthopedic attending physician (25.45 ± 1.92 min) and the radiology attending physician (26.21 ± 1.20 min), with a statistically significant difference (p < 0.05).

CONCLUSION: The model based on the integration of blockchain technology and AI can realize safe, collaborative, and convenient assisted diagnosis of TPF. Through the aggregation of training parameters by decentralized algorithms, it can achieve model construction without data leaving the hospital and may exert clinical application value in the emergency settings.

PMID:39984717 | DOI:10.1007/s00068-025-02793-y

Langenbecks Arch Surg. 2025 Feb 22;410(1):80. doi: 10.1007/s00423-025-03641-8.

ABSTRACT

PURPOSE: Several surgical-skill assessment tools emphasize the importance of efficient tissue-dissection, whose assessment relies on human judgment and is thus subject to bias. Automated assessment may help solve this problem. This study aimed to verify the feasibility of surgical-skill assessment using a deep learning-based recognition model.

METHODS: This retrospective study used multicenter intraoperative videos of laparoscopic colorectal surgery (sigmoidectomy or high anterior resection) for colorectal cancer obtained from 766 cases across Japan. Three groups with different skill levels were distinguished: high-, intermediate-, and low-skill. We developed a model to recognize tissue dissection by the monopolar device using deep learning-based computer-vision technology. Tissue-dissection time per monopolar device appearance time (efficient-dissection time ratio) was extracted as a quantitative parameter describing efficient dissection. We automatically measured the efficient-dissection time ratio using the recognition model of 8 surgical instruments and tissue-dissection on/off classification model. The efficient-dissection time ratio was compared among groups; the feasibility of distinguishing them was explored using the model. The model-calculated parameters were evaluated to determine whether they could differentiate high-, intermediate-, and low-skill groups.

RESULTS: The tissue-dissection recognition model had an overall accuracy of 0.91. There was a moderate correlation (0.542; 95% confidence interval, 0.288-0.724; P < 0.001) between manually and automatically measured efficient-dissection time ratios. Efficient-dissection time ratios by this model were significantly higher in the high-skill than in intermediate-skill (P = 0.0081) and low-skill (P = 0.0249) groups.

CONCLUSION: An automated efficient-dissection assessment model using a monopolar device was constructed with a feasible automated skill-assessment method.

PMID:39984705 | DOI:10.1007/s00423-025-03641-8

Sci Rep. 2025 Feb 21;15(1):6362. doi: 10.1038/s41598-025-90185-y.

ABSTRACT

X-ray inspection is a crucial technique for identifying defects in castings, capable of revealing minute internal flaws such as pores and inclusions. However, traditional methods rely on the subjective judgment of experts, are time-consuming, and prone to errors, which negatively impact the efficiency and accuracy of inspections. Therefore, the development of an automated defect detection model is of significant importance for enhancing the scientific rigor and precision of casting inspections. In this study, we propose a deep learning model specifically designed for detecting small-scale defects in castings. The model employs an end-to-end network architecture and features a loss function based on the Wasserstein distance, which is tailored to optimize the training process for small defect targets, thereby improving detection accuracy. Additionally, we have innovatively developed a dual-layer Encoder-Decoder multi-scale feature extraction architecture, BiSDE, based on the Hadamard product, aimed at enhancing the model's ability to recognize and locate small targets. To evaluate the performance of the proposed model, we conducted a series of experiments, including comparative tests with current state-of-the-art object detection models such as Yolov9, FasterNet, Yolov8, and Detr, as well as ablation studies on the model's components. The results demonstrate that our model achieves at least a 5.3% improvement in Mean Average Precision (MAP) over the existing state-of-the-art models. Furthermore, the inclusion of each component significantly enhanced the overall performance of the model. In conclusion, our research not only validates the effectiveness of the proposed small-scale defect detection model in improving detection precision but also offers broad prospects for the automation and intelligent development of industrial defect inspection.

PMID:39984609 | DOI:10.1038/s41598-025-90185-y

Sci Rep. 2025 Feb 21;15(1):6312. doi: 10.1038/s41598-025-90024-0.

ABSTRACT

Successful translation of artificial intelligence (AI) models into clinical practice, across clinical domains, is frequently hindered by the lack of image quality control. Diagnostic models are often trained on images with no denotation of image quality in the training data; this, in turn, can lead to misclassifications by these models when implemented in the clinical setting. In the case of cervical images, quality classification is a crucial task to ensure accurate detection of precancerous lesions or cancer; this is true for both gynecologic-oncologists' (manual) and diagnostic AI models' (automated) predictions. Factors that impact the quality of a cervical image include but are not limited to blur, poor focus, poor light, noise, obscured view of the cervix due to mucus and/or blood, improper position, and over- and/or under-exposure. Utilizing a multi-level image quality ground truth denoted by providers, we generated an image quality classifier following a multi-stage model selection process that investigated several key design choices on a multi-heterogenous "SEED" dataset of 40,534 images. We subsequently validated the best model on an external dataset ("EXT"), comprising 1,340 images captured using a different device and acquired in different geographies from "SEED". We assessed the relative impact of various axes of data heterogeneity, including device, geography, and ground-truth rater on model performance. Our best performing model achieved an area under the receiver operating characteristics curve (AUROC) of 0.92 (low quality, LQ vs. rest) and 0.93 (high quality, HQ vs. rest), and a minimal total %extreme misclassification (%EM) of 2.8% on the internal validation set. Our model also generalized well externally, achieving corresponding AUROCs of 0.83 and 0.82, and %EM of 3.9% when tested out-of-the-box on the external validation ("EXT") set. Additionally, our model was geography agnostic with no meaningful difference in performance across geographies, did not exhibit catastrophic forgetting upon retraining with new data, and mimicked the overall/average ground truth rater behavior well. Our work represents one of the first efforts at generating and externally validating an image quality classifier across multiple axes of data heterogeneity to aid in visual diagnosis of cervical precancer and cancer. We hope that this will motivate the accompaniment of adequate guardrails for AI-based pipelines to account for image quality and generalizability concerns.

PMID:39984572 | DOI:10.1038/s41598-025-90024-0

Biomed Phys Eng Express. 2025 Feb 21. doi: 10.1088/2057-1976/adb8f1. Online ahead of print.

ABSTRACT

OBJECTIVE: To determine the minimum detector width along the fan-angle direction in offset-detector cone-beam breast CT for multiple advanced reconstruction algorithms and to investigate the effect on radiation dose, scatter, and image quality.

APPROACH: Complete sinograms (m × n = 1024 × 768 pixels) of 30 clinical breast CT datasets previously acquired on a clinical-prototype cone-beam breast CT system were reconstructed using Feldkamp-Davis-Kress (FDK) algorithm and served as the reference. Complete sinograms were retrospectively truncated to varying widths to understand the limits of four image reconstruction algorithms - FDK with redundancy weighting (FDK-W), compressed-sensing based FRIST, fully-supervised MS-RDN, and self-supervised AFN. Upon determining the truncation limits, numerical phantoms generated by segmenting the reference reconstructions into skin, adipose, and fibroglandular tissues were used to determine the radiation dose and scatter-to-primary ratio (SPR) using Monte Carlo simulations.

MAIN RESULTS: FDK-W, FRIST, and MS-RDN showed artifacts when m < 596, whereas AFN reconstructed images without artifacts for m>=536. Reducing the detector width reduced signal-difference to noise ratio (SDNR) for FDK-W, whereas FRIST, MS-RDN and AFN maintained or improved SDNR. Reference reconstruction and AFN with m=536 had similar quantitative measures of image quality.

SIGNIFICANCE: For the 30 cases, AFN with m=536 reduced the radiation dose and SPR by 37.85% and 33.46%, respectively, compared to the reference. Qualitative and quantitative image quality indicate the feasibility of AFN for offset-detector cone-beam breast CT. Radiation dose and SPR were simultaneously reduced with a 536 ×768 detector and when used in conjunction with AFN algorithm had similar image quality as the reference reconstruction.

PMID:39983239 | DOI:10.1088/2057-1976/adb8f1

J Neural Eng. 2025 Feb 21. doi: 10.1088/1741-2552/adb90c. Online ahead of print.

ABSTRACT

OBJECTIVE: Humans spend a significant portion of their lives in sleep (an essential driver of body metabolism). Moreover, as sleep deprivation could cause various health complications, it is crucial to develop an automatic sleep stage detection model to facilitate the tedious manual labeling process. Notably, recently proposed sleep staging algorithms lack model explainability and still require performance improvement.

APPROACH: We implemented multiscale neurophysiology-mimicking kernels to capture sleep-related electroencephalogram (EEG) activities at varying frequencies and temporal lengths; the implemented model was named "Multiscale Temporal Convolutional Neural Network (MTCNN)." Further, we evaluated its performance using an open-source dataset (Sleep-EDF Database Expanded comprising 153 days of polysomnogram data).

MAIN RESULTS: By investigating the learned kernel weights, we observed that MTCNN detected the EEG activities specific to each sleep stage, such as the frequencies, K-complexes, and sawtooth waves. Furthermore, regarding the characterization of these neurophysiologically significant features, MTCNN demonstrated an overall accuracy (OAcc) of 91.12% and a Cohen kappa coefficient of 0.86 in the cross-subject paradigm. Notably, it demonstrated an OAcc of 88.24% and a Cohen kappa coefficient of 0.80 in the leave-few-days-out analysis. Our MTCNN model also outperformed the existing deep learning models in sleep stage classification even when it was trained with only 16% of the total EEG data, achieving an OAcc of 85.62% and a Cohen kappa coefficient of 0.75 on the remaining 84% of testing data.

SIGNIFICANCE: The proposed MTCNN enables model explainability and it can be trained with lesser amount of data, which is beneficial to its application in the real-world because large amounts of training data are not often and readily available.

PMID:39983236 | DOI:10.1088/1741-2552/adb90c

J Chem Inf Model. 2025 Feb 21. doi: 10.1021/acs.jcim.4c01371. Online ahead of print.

ABSTRACT

The importance of drug toxicity assessment lies in ensuring the safety and efficacy of the pharmaceutical compounds. Predicting toxicity is crucial in drug development and risk assessment. This study compares the performance of GPT-4 and GPT-4o with traditional deep-learning and machine-learning models, WeaveGNN, MorganFP-MLP, SVC, and KNN, in predicting molecular toxicity, focusing on bone, neuro, and reproductive toxicity. The results indicate that GPT-4 is comparable to deep-learning and machine-learning models in certain areas. We utilized GPT-4 combined with molecular docking techniques to study the cardiotoxicity of three specific targets, examining traditional Chinese medicinal materials listed as both food and medicine. This approach aimed to explore the potential cardiotoxicity and mechanisms of action. The study found that components in Black Sesame, Ginger, Perilla, Sichuan Pagoda Tree Fruit, Galangal, Turmeric, Licorice, Chinese Yam, Amla, and Nutmeg exhibit toxic effects on cardiac target Cav1.2. The docking results indicated significant binding affinities, supporting the hypothesis of potential cardiotoxic effects.This research highlights the potential of ChatGPT in predicting molecular properties and its significance in medicinal chemistry, demonstrating its facilitation of a new research paradigm: with a data set, high-accuracy learning models can be generated without requiring computational knowledge or coding skills, making it accessible and easy to use.

PMID:39982968 | DOI:10.1021/acs.jcim.4c01371

PLoS One. 2025 Feb 21;20(2):e0316989. doi: 10.1371/journal.pone.0316989. eCollection 2025.

ABSTRACT

Hyperpartisan news consists of articles with strong biases that support specific political parties. The spread of such news increases polarization among readers, which threatens social unity and democratic stability. Automated tools can help identify hyperpartisan news in the daily flood of articles, offering a way to tackle these problems. With recent advances in machine learning and deep learning, there are now more methods available to address this issue. This literature review collects and organizes the different methods used in previous studies on hyperpartisan news detection. Using the PRISMA methodology, we reviewed and systematized approaches and datasets from 81 articles published from January 2015 to 2024. Our analysis includes several steps: differentiating hyperpartisan news detection from similar tasks, identifying text sources, labeling methods, and evaluating models. We found some key gaps: there is no clear definition of hyperpartisanship in Computer Science, and most datasets are in English, highlighting the need for more datasets in minority languages. Moreover, the tendency is that deep learning models perform better than traditional machine learning, but Large Language Models' (LLMs) capacities in this domain have been limitedly studied. This paper is the first to systematically review hyperpartisan news detection, laying a solid groundwork for future research.

PMID:39982955 | DOI:10.1371/journal.pone.0316989

Ophthalmol Ther. 2025 Feb 21. doi: 10.1007/s40123-025-01109-y. Online ahead of print.

ABSTRACT

Fundus fluorescein angiography (FFA) serves as the current gold standard for visualizing retinal vasculature and detecting various fundus diseases, but its interpretation is labor-intensive and requires much expertise from ophthalmologists. The medical application of artificial intelligence (AI), especially deep learning and machine learning, has revolutionized the field of automatic FFA image analysis, leading to the rapid advancements in AI-assisted lesion detection, diagnosis, and report generation. This review examined studies in PubMed, Web of Science, and Google Scholar databases from January 2019 to August 2024, with a total of 23 articles incorporated. By integrating current research findings, this review highlights crucial breakthroughs in AI-assisted FFA analysis and explores their potential implications for ophthalmic clinical practice. These advances in AI-assisted FFA analysis have shown promising results in improving diagnostic accuracy and workflow efficiency. However, further research is needed to enhance model transparency and ensure robust performance across diverse populations. Challenges such as data privacy and technical infrastructure remain for broader clinical applications.

PMID:39982648 | DOI:10.1007/s40123-025-01109-y

Ann Hematol. 2025 Feb 21. doi: 10.1007/s00277-025-06254-9. Online ahead of print.

ABSTRACT

Monitoring anemia recovery is crucial for clinical intervention. Morphological assessment of red blood cells (RBCs) with peripheral blood smears (PBSs) provides additional information beyond routine blood tests. However, the PBS test is labor-intensive, reliant on manual analysis, and susceptible to variability in expert interpretations. Here we introduce a deep semi-supervised learning method, RBCMatch, to classify RBCs during anemia recovery. Using an acute hemolytic anemic mouse model, PBS images at four different time points during anemia recovery were acquired and segmented into 10,091 single RBC images, with only 5% annotated and used in model training. By employing the semi-supervised strategy Fixmatch, RBCMatch achieved an impressive average classification accuracy of 91.2% on the validation dataset and 87.5% on a held-out dataset, demonstrating its superior accuracy and robustness compared to supervised learning methods, especially when labeled samples are scarce. To characterize the anemia recovery process, principal components (PCs) of RBC embeddings were extracted and visualized. Our results indicated that RBC embeddings quantified the state of anemia recovery, and the second PC had a strong correlation with RBC count and hemoglobin concentration, demonstrating the model's ability to accurately depict RBC morphological changes during anemia recovery. Thus, this study provides a valuable tool for the automatic classification of RBCs and offers novel insights into the assessment of anemia recovery, with the potential to aid in clinical decision-making and prognosis analysis in the future.

PMID:39982510 | DOI:10.1007/s00277-025-06254-9

Radiologie (Heidelb). 2025 Feb 21. doi: 10.1007/s00117-025-01416-2. Online ahead of print.

ABSTRACT

BACKGROUND: Given the increasing number of radiological examinations, large language models (LLMs) offer promising support in radiology. Optimized interaction is essential to ensure reliable results.

OBJECTIVES: This article provides an overview of interaction techniques such as prompt engineering, zero-shot learning, and retrieval-augmented generation (RAG) and gives practical tips for their application in radiology.

MATERIALS AND METHODS: Demonstration of interaction techniques based on practical examples with concrete recommendations for their application in routine radiological practice.

RESULTS: Advanced interaction techniques allow task-specific adaptation of LLMs without the need for retraining. The creation of precise prompts and the use of zero-shot and few-shot learning can significantly improve response quality. RAG enables the integration of current and domain-specific information into LLM tools, increasing the accuracy and relevance of the generated content.

CONCLUSIONS: The use of prompt engineering, zero-shot and few-shot learning, and RAG can optimize interaction with LLMs in radiology. Through these targeted strategies, radiologists can efficiently integrate general chatbots into routine practice to improve patient care.

PMID:39982460 | DOI:10.1007/s00117-025-01416-2

Proteomes. 2025 Feb 17;13(1):10. doi: 10.3390/proteomes13010010.

ABSTRACT

Background: The interaction network between the human host proteins and the proteins of the gut bacteria is essential for the establishment of human health, and its dysregulation directly contributes to disease development. Despite its great importance, experimental data on protein-protein interactions (PPIs) between these species are sparse due to experimental limitations. Methods: This study presents a deep learning-based framework for predicting PPIs between human and gut bacterial proteins using structural data. The framework leverages graph-based protein representations and variational autoencoders (VAEs) to extract structural embeddings from protein graphs, which are then fused through a Bi-directional Cross-Attention module to predict interactions. The model addresses common challenges in PPI datasets, such as class imbalance, using focal loss to emphasize harder-to-classify samples. Results: The results demonstrated that this framework exhibits robust performance, with high precision and recall across validation and test datasets, underscoring its generalizability. By incorporating proteoforms in the analysis, the model accounts for the structural complexity within proteomes, making predictions biologically relevant. Conclusions: These findings offer a scalable tool for investigating the interactions between the host and the gut microbiota, potentially yielding new treatment targets and diagnostics for disorders linked to the microbiome.

PMID:39982320 | DOI:10.3390/proteomes13010010