Radiat Oncol. 2025 May 19;20(1):80. doi: 10.1186/s13014-025-02634-7.

ABSTRACT

PURPOSE: Accurate pre-treatment dose prediction is essential for efficient radiotherapy planning. Although deep learning models have advanced automated dose distribution, comprehensive multi-tumor analyses remain scarce. This study assesses deep learning models for dose prediction across diverse tumor types, combining objective and subjective evaluation methods.

METHODS AND MATERIALS: We included 622 patients with planning data across various tumor sites: nasopharyngeal carcinoma (n = 29), esophageal carcinoma (n = 82), left-sided breast carcinoma (n = 107), right-sided breast carcinoma (n = 95), cervical carcinoma treated with radical radiotherapy (n = 84), postoperative cervical carcinoma (n = 122), and rectal carcinoma (n = 103). Dose predictions were generated using U-Net, Flex-Net, and Highres-Net models, with data split into training (60%), validation (20%), and testing (20%) sets. Quantitative comparisons used normalized dose difference (NDD) and dose-volume histogram (DVH) metrics, and qualitative assessments by radiation oncologists were performed on the testing set.

RESULTS: Predicted and clinical doses correlated well, with NDD values under 3% for tumor targets in nasopharyngeal, breast, and postoperative cervical cancer. Qualitative assessments revealed that U-Net, Flex-Net, and Highres-Net achieved the highest accuracy in cervical radical, breast/rectal/postoperative cervical, and nasopharyngeal/esophageal cancers, respectively. Among the test cases (n = 123), 53.7% were deemed clinically acceptable and 32.5% required minor adjustments. The "Best Selection" approach, combining strengths of all three models, raised clinical acceptance to 62.6%.

CONCLUSION: This study demonstrates that automated dose prediction can provide a robust starting point for rapid plan generation. Leveraging model-specific strengths through the "Best Selection" approach enhances prediction accuracy and shows potential to improve clinical efficiency across multiple tumor types.

PMID:40390053 | DOI:10.1186/s13014-025-02634-7

BMC Med Imaging. 2025 May 19;25(1):173. doi: 10.1186/s12880-025-01692-3.

ABSTRACT

OBJECTIVES: To establish and validate deep learning (DL) models based on pre-treatment multiparametric magnetic resonance imaging (MRI) images of primary rectal cancer and basic clinical data for the prediction of synchronous liver metastases (SLM) in patients with Rectal cancer (RC).

METHODS: In this retrospective study, 176 and 31 patients with RC who underwent multiparametric MRI from two centers were enrolled in the primary and external validation cohorts, respectively. Clinical factors, including sex, primary tumor site, CEA level, and CA199 level were assessed. A clinical feature (CF) model was first developed by multivariate logistic regression, then two residual network DL models were constructed based on multiparametric MRI of primary cancer with or without CF incorporation. Finally, the SLM prediction models were validated by 5-fold cross-validation and external validation. The performance of the models was evaluated by decision curve analysis (DCA) and receiver operating characteristic (ROC) analysis.

RESULTS: Among three SLM prediction models, the Combined DL model integrating primary tumor MRI and basic clinical data achieved the best performance (AUC = 0.887 in primary study cohort; AUC = 0.876 in the external validation cohort). In the primary study cohort, the CF model, MRI DL model, and Combined DL model achieved AUCs of 0.816 (95% CI: 0.750, 0.881), 0.788 (95% CI: 0.720, 0.857), and 0.887 (95% CI: 0.834, 0.940) respectively. In the external validation cohort, the CF model, DL model without CF, and DL model with CF achieved AUCs of 0.824 (95% CI: 0.664, 0.984), 0.662 (95% CI: 0.461, 0.863), and 0.876 (95% CI: 0.728, 1.000), respectively.

CONCLUSION: The combined DL model demonstrates promising potential to predict SLM in patients with RC, thereby making individualized imaging test strategies.

CLINICAL RELEVANCE STATEMENT: Accurate synchronous liver metastasis (SLM) risk stratification is important for treatment planning and prognosis improvement. The proposed DL signature may be employed to better understand an individual patient's SLM risk, aiding in treatment planning and selection of further imaging examinations to personalize clinical decisions.

CLINICAL TRIAL NUMBER: Not applicable.

PMID:40389920 | DOI:10.1186/s12880-025-01692-3

BMC Med Imaging. 2025 May 19;25(1):171. doi: 10.1186/s12880-025-01703-3.

ABSTRACT

OBJECTIVE: Development of a deep learning model for accurate preoperative identification of glioblastoma and solitary brain metastases by combining multi-centre and multi-sequence magnetic resonance images and comparison of the performance of different deep learning models.

METHODS: Clinical data and MR images of a total of 236 patients with pathologically confirmed glioblastoma and single brain metastases were retrospectively collected from January 2019 to May 2024 at Provincial Hospital of Shandong First Medical University, and the data were randomly divided into a training set and a test set according to the ratio of 8:2, in which the training set contained 197 cases and the test set contained 39 cases; the images were preprocessed and labeled with the tumor regions. The images were pre-processed and labeled with tumor regions, and different MRI sequences were input individually or in combination to train the deep learning model 3D ResNet-18, and the optimal sequence combinations were obtained by five-fold cross-validation enhancement of the data inputs and training of the deep learning models 3D Vision Transformer (3D Vit), 3D DenseNet, and 3D VGG; the working characteristic curves (ROCs) of subjects were plotted, and the area under the curve (AUC) was calculated. The area under the curve (AUC), accuracy, precision, recall and F1 score were used to evaluate the discriminative performance of the models. In addition, 48 patients with glioblastoma and single brain metastases from January 2020 to December 2022 were collected from the Affiliated Cancer Hospital of Shandong First Medical University as an external test set to compare the discriminative performance, robustness and generalization ability of the four deep learning models.

RESULTS: In the comparison of the discriminative effect of different MRI sequences, the three sequence combinations of T1-CE, T2, and T2-Flair gained discriminative effect, with the accuracy and AUC values of 0.8718 and 0.9305, respectively; after the four deep learning models were inputted into the aforementioned sequence combinations, the accuracy and AUC of the external validation of the 3D ResNet-18 model were 0.8125, respectively, 0.8899, all of which are the highest among all models.

CONCLUSIONS: A combination of multi-sequence MR images and a deep learning model can efficiently identify glioblastoma and solitary brain metastases preoperatively, and the deep learning model 3D ResNet-18 has the highest efficacy in identifying the two types of tumours.

PMID:40389875 | DOI:10.1186/s12880-025-01703-3

Clin Rheumatol. 2025 May 19. doi: 10.1007/s10067-025-07481-1. Online ahead of print.

ABSTRACT

OBJECTIVE: To develop and validate a deep learning radiomics fusion model (DLR) based on ultrasound (US) images to identify bone erosion in rheumatoid arthritis (RA) patients.

METHODS: A total of 432 patients with RA at two institutions were collected. Three hundred twelve patients from center 1 were randomly divided into a training set (N = 218) and an internal test set (N = 94) in a 7:3 ratio; meanwhile, 124 patients from center 2 were as an external test set. Radiomics (Rad) and deep learning (DL) features were extracted based on hand-crafted radiomics and deep transfer learning networks. The least absolute shrinkage and selection operator regression was employed to establish DLR fusion feature from the Rad and DL features. Subsequently, 10 machine learning algorithms were used to construct models and the final optimal model was selected. The performance of models was evaluated using receiver operating characteristic (ROC) and decision curve analysis (DCA). The diagnostic efficacy of sonographers was compared with and without the assistance of the optimal model.

RESULTS: LR was chosen as the optimal algorithm for model construction account for superior performance (Rad/DL/DLR: area under the curve [AUC] = 0.906/0.974/0.979) in the training set. In the internal test set, DLR_LR as the final model had the highest AUC (AUC = 0.966), which was also validated in the external test set (AUC = 0.932). With the aid of DLR_LR model, the overall performance of both junior and senior sonographers improved significantly (P < 0.05), and there was no significant difference between the junior sonographer with DLR_LR model assistance and the senior sonographer without assistance (P > 0.05).

CONCLUSION: DLR model based on US images is the best performer and is expected to become an important tool for identifying bone erosion in RA patients. Key Points • DLR model based on US images is the best performer in identifying BE in RA patients. • DLR model may assist the sonographers to improve the accuracy of BE evaluations.

PMID:40389785 | DOI:10.1007/s10067-025-07481-1

Aesthetic Plast Surg. 2025 May 19. doi: 10.1007/s00266-025-04954-1. Online ahead of print.

ABSTRACT

The integration of artificial intelligence (AI) into rhinoplasty has transformed preoperative planning and patient communication by providing highly accurate simulations of postoperative outcomes. AI-driven models, including deep learning and generative adversarial networks (GANs), have demonstrated the ability to predict nasal shapes, learn surgical styles, and refine aesthetic planning. However, despite these advancements, AI remains an imperfect predictor of surgical results, as it cannot account for individual healing processes, tissue behavior, and long-term nasal remodeling. Moreover, ethical concerns arise regarding bias in AI-generated predictions, the reinforcement of unattainable beauty standards, and the psychological impact on patients, particularly in the era of social media-driven aesthetics. Additionally, issues related to data privacy and the medico-legal risks associated with unrealistic patient expectations must be addressed. While AI is a powerful tool for enhancing rhinoplasty planning, it should be considered an adjunct rather than a replacement for surgical expertise. Future developments must refine AI models to incorporate patient-specific variables while maintaining a balanced approach that prioritizes ethical medical practice and individualized patient care.Level of Evidence V This journal requires that authors assign a level of evidence to each article. For a full description of these Evidence-Based Medicine ratings, please refer to the Table of Contents or the online Instructions to Authors www.springer.com/00266 .

PMID:40389738 | DOI:10.1007/s00266-025-04954-1

Sci Rep. 2025 May 19;15(1):17306. doi: 10.1038/s41598-025-99990-x.

ABSTRACT

Heart disease is one of the major causes of death worldwide, and the traditional diagnostic procedures typically cause delays in treatment, particularly in low-resource regions. In this article, we propose a novel IoT-based Quantum Kernel-Enhanced Sardine Diffusion Attention Network (Qua-KSar-DCK-ArNet) for real-time prediction of heart disease. The system is capable of continuously monitoring heart-related data such as ECG and heart rate via IoT sensors. Quantum Clustering with k-Means is applied to cluster the data, and Z-score Min-Max Normalization is applied for preprocessing. Fast Point Transformer is utilized to identify salient features. The Qua-KSar-DCK-ArNet model, a combination of quantum and classical deep learning methods, classifies the data for predicting the risk of heart disease. The system is fast and accurate, with an accuracy of 99%, significantly improving patient outcomes, especially in resource-scarce regions.

PMID:40389708 | DOI:10.1038/s41598-025-99990-x

Sci Rep. 2025 May 19;15(1):17331. doi: 10.1038/s41598-025-02018-7.

ABSTRACT

Dementia spectrum disorders, characterized by progressive cognitive decline, pose a significant global health burden. Early screening and diagnosis are essential for timely and accurate treatment, improving patient outcomes and quality of life. This study investigated dynamic features of resting-state electroencephalography (EEG) functional connectivity to identify characteristic patterns of dementia subtypes, such as Alzheimer's disease (AD) and frontotemporal dementia (FD), and to evaluate their potential as biomarkers. We extracted distinctive statistical features, including mean, variance, skewness, and Shannon entropy, from brain connectivity measures, revealing common alterations in dementia, specifically a generalized disruption of Alpha-band connectivity. Distinctive characteristics were found, including generalized Delta-band hyperconnectivity with increased complexity in AD and disrupted phase-based connectivity in Theta, Beta, and Gamma bands for FD. We also employed a convolutional neural network model, enhanced with these dynamic features, to differentiate between dementia subtypes. Our classification models achieved a multiclass classification accuracy of 93.6% across AD, FD, and healthy control groups. Furthermore, the model demonstrated 97.8% and 96.7% accuracy in differentiating AD and FD from healthy controls, respectively, and 97.4% accuracy in classifying AD and FD in pairwise classification. These establish a high-performance deep learning framework utilizing dynamic EEG connectivity patterns as potential biomarkers, offering a promising approach for early screening and diagnosis of dementia spectrum disorders using EEG. Our findings suggest that analyzing brain connectivity dynamics as a network and during cognitive tasks could offer more valuable information for diagnosis, assessing disease severity, and potentially identifying personalized neurological deficits.

PMID:40389648 | DOI:10.1038/s41598-025-02018-7

Sci Rep. 2025 May 19;15(1):17372. doi: 10.1038/s41598-025-87529-z.

ABSTRACT

Deep learning models are inherently vulnerable to adversarial examples, particularly in black-box settings where attackers have limited knowledge of the target model. Existing attack algorithms often face challenges in balancing effectiveness and efficiency. Adversarial perturbations generated in such settings can be suboptimal and require large query budgets to achieve high success rates. In this paper, we investigate the transferability of Multiple Universal Adversarial Perturbations (MUAPs), showing that they can affect a large portion of samples across different models. Based on this insight, we propose SMPack, a staged black-box adversarial example generation algorithm that integrates surrogate and query schemes. By combining MUAPs with surrogate models, SMPack effectively overcomes the black-box constraints and improves the efficiency of generating adversarial examples. Additionally, we optimize this process using a Genetic Algorithm (GA), allowing for efficient search of the perturbation space while conserving query budget. We evaluated SMPack against eight popular attack algorithms: OnePixel, SimBA, FNS, GA, SFGSM, SPGD, FGSM, and PGD, using four publicly available datasets: MNIST, SVHN, CIFAR-10, and ImageNet. The experiments involved 500 random correctly classified samples for each dataset. Our results show that SMPack outperforms existing black-box attack methods in both attack success rate (ASR) and query efficiency, while maintaining competitive performance with white-box methods. SMPack provides an efficient and effective solution for generating adversarial examples in black-box settings. The integration of MUAPs, surrogate schemes, and genetic optimization addresses the key limitations of existing methods, offering a robust alternative for generating adversarial perturbations with reduced query budget.

PMID:40389546 | DOI:10.1038/s41598-025-87529-z

Sci Rep. 2025 May 19;15(1):17363. doi: 10.1038/s41598-025-01758-w.

ABSTRACT

Conventional image formats have limited information conveyance, while Hyperspectral Imaging (HSI) offers a broader representation through continuous spectral bands, capturing hundreds of spectral features. However, this abundance leads to redundant information, posing a computational challenge for deep learning models. Thus, models must effectively extract indicative features. HSI's non-linear nature, influenced by environmental factors, necessitates both linear and non-linear modeling techniques for feature extraction. While PCA and ICA, being linear methods, may overlook complex patterns, Autoencoders (AE) can capture and represent non-linear features. Yet, AEs can be biased by unbalanced datasets, emphasizing majority class features and neglecting minority class characteristics, highlighting the need for careful dataset preparation. To address this, the Dual-Path AE (D-Path-AE) model has been proposed, which enhances non-linear feature acquisition through concurrent encoding pathways. This model also employs a down-sampling strategy to reduce bias towards majority classes. The study compared the efficacy of dimensionality reduction using the Naïve Autoencoder (Naïve AE) and D-Path-AE. Classification capabilities were assessed using Decision Tree, Support Vector Machine, and K-Nearest Neighbors (KNN) classifiers on datasets from Pavia Center, Salinas, and Kennedy Space Center. Results demonstrate that the D-Path-AE outperforms both linear dimensionality reduction models and Naïve AE, achieving an Overall Accuracy of up to 98.31% on the Pavia Center dataset using the KNN classifier, indicating superior classification capabilities.

PMID:40389536 | DOI:10.1038/s41598-025-01758-w

Rhinology. 2025 May 19. doi: 10.4193/Rhin25.044. Online ahead of print.

ABSTRACT

BACKGROUND: Sinonasal pathology can be complex and requires a systematic and meticulous approach. Artificial Intelligence (AI) has the potential to improve diagnostic accuracy and efficiency in sinonasal imaging, but its clinical applicability remains an area of ongoing research. This systematic review evaluates the methodologies and clinical relevance of AI in detecting sinonasal pathology through radiological imaging.

METHODOLOGY: Key search terms included "artificial intelligence," "deep learning," "machine learning," "neural network," and "paranasal sinuses,". Abstract and full-text screening was conducted using predefined inclusion and exclusion criteria. Data were extracted on study design, AI architectures used (e.g., Convolutional Neural Networks (CNN), Machine Learning classifiers), and clinical characteristics, such as imaging modality (e.g., Computed Tomography (CT), Magnetic Resonance Imaging (MRI)).

RESULTS: A total of 53 studies were analyzed, with 85% retrospective, 68% single-center, and 92.5% using internal databases. CT was the most common imaging modality (60.4%), and chronic rhinosinusitis without nasal polyposis (CRSsNP) was the most studied condition (34.0%). Forty-one studies employed neural networks, with classification as the most frequent AI task (35.8%). Key performance metrics included Area Under the Curve (AUC), accuracy, sensitivity, specificity, precision, and F1-score. Quality assessment based on CONSORT-AI yielded a mean score of 16.0 ± 2.

CONCLUSIONS: AI shows promise in improving sinonasal imaging interpretation. However, as existing research is predominantly retrospective and single-center, further studies are needed to evaluate AI's generalizability and applicability. More research is also required to explore AI's role in treatment planning and post-treatment prediction for clinical integration.

PMID:40388840 | DOI:10.4193/Rhin25.044

Proc Natl Acad Sci U S A. 2025 May 27;122(21):e2412261122. doi: 10.1073/pnas.2412261122. Epub 2025 May 19.

ABSTRACT

Enhancing the reliability and reproducibility of optical microscopy by reducing specimen irradiance continues to be an important biotechnology target. As irradiance levels are reduced, however, the particle nature of light is heightened, giving rise to Poisson noise, or photon sparsity that restricts only a few (0.5%) image pixels to comprise a photon. Photon sparsity can be addressed by collecting approximately 200 photons per pixel; this, however, requires long acquisitions and, as such, suboptimal imaging rates. Here, we introduce near-zero photon bioimaging, a method that operates at kHz rates and 10,000-fold lower irradiance than standard microscopy. To achieve this level of performance, we uniquely combined a judiciously designed epifluorescence microscope enabling ultralow background levels and AI that learns to reconstruct biological images from as low as 0.01 photons per pixel. We demonstrate that near-zero photon bioimaging captures the structure of multicellular and subcellular features with high fidelity, including features represented by nearly zero photons. Beyond optical microscopy, the near-zero photon bioimaging paradigm can be applied in remote sensing, covert applications, and biomedical imaging that utilize damaging or quantum light.

PMID:40388622 | DOI:10.1073/pnas.2412261122

PLoS One. 2025 May 19;20(5):e0310230. doi: 10.1371/journal.pone.0310230. eCollection 2025.

ABSTRACT

The rapid growth of Android applications has led to an increase in security threats, while traditional detection methods struggle to combat advanced malware, such as polymorphic and metamorphic variants. To address these challenges, this study introduces a hybrid deep learning model (DBN-GRU) that integrates Deep Belief Networks (DBN) for static analysis and Gated Recurrent Units (GRU) for dynamic behavior modeling to enhance malware detection accuracy and efficiency. The model extracts static features (permissions, API calls, intent filters) and dynamic features (system calls, network activity, inter-process communication) from Android APKs, enabling a comprehensive analysis of application behavior.The proposed model was trained and tested on the Drebin dataset, which includes 129,013 applications (5,560 malware and 123,453 benign).Performance evaluation against NMLA-AMDCEF, MalVulDroid, and LinRegDroid demonstrated that DBN-GRU achieved 98.7% accuracy, 98.5% precision, 98.9% recall, and an AUC of 0.99, outperforming conventional models.In addition, it exhibits faster preprocessing, feature extraction, and malware classification times, making it suitable for real-time deployment.By bridging static and dynamic detection methodologies, the DBN-GRU enhances malware detection capabilities while reducing false positives and computational overhead.These findings confirm the applicability of the proposed model in real-world Android security applications, offering a scalable and high-performance malware detection solution.

PMID:40388500 | DOI:10.1371/journal.pone.0310230

PLoS One. 2025 May 19;20(5):e0313692. doi: 10.1371/journal.pone.0313692. eCollection 2025.

ABSTRACT

Surveillance systems are integral to ensuring public safety by detecting unusual incidents, yet existing methods often struggle with accuracy and robustness. This study introduces an advanced framework for anomaly recognition in surveillance, leveraging deep learning to address these challenges and achieve significant improvements over current techniques. The framework begins with preprocessing input images using histogram equalization to enhance feature visibility. It then employs two DCNNs for feature extraction: a novel 63-layer CNN, "Up-to-the-Minute-Net," and the established Inception-Resnet-v2. The features extracted by both models are fused and optimized through two sophisticated feature selection techniques: Dragonfly and Genetic Algorithm (GA). The optimization process involves rigorous experimentation with 5- and 10-fold cross-validation to evaluate performance across various feature sets. The proposed approach achieves an unprecedented 99.9% accuracy in 5-fold cross-validation using the GA optimizer with 2500 selected features, demonstrating a substantial leap in accuracy compared to existing methods. This study's contribution lies in its innovative combination of deep learning models and advanced feature optimization techniques, setting a new benchmark in the field of anomaly recognition for surveillance systems and showcasing the potential for practical real-world applications.

PMID:40388481 | DOI:10.1371/journal.pone.0313692

PLoS One. 2025 May 19;20(5):e0324047. doi: 10.1371/journal.pone.0324047. eCollection 2025.

ABSTRACT

This study aims to develop a predictive hybrid model for a grid-connected PV system with DC-DC optimizers, designed to operate in extreme altitude conditions at 3800 m above sea level. This approach seeks to address the "curse of dimensionality" by reducing model complexity and improving its accuracy by combining the recursive feature removal (RFE) method with advanced regularization techniques, such as Lasso, Ridge, and Bayesian Ridge. The research used a photovoltaic system composed of monocrystalline modules, DC-DC optimizers and a 3000 W inverter. The data obtained from the system were divided into training and test sets, where RFE identified the most relevant variables, eliminating the reactive power of AC. Subsequently, the three regularization models were trained with these selected variables and evaluated using metrics such as precision, mean absolute error, mean square error and coefficient of determination. The results showed that RFE - Bayesian Ridge obtained the highest accuracy (0.999935), followed by RFE - Ridge, while RFE - Lasso had a slightly lower performance and also obtained an exceptionally low MASE (0.0034 for Bayesian and Ridge, compared to 0.0065 for Lasso). All models complied with the necessary statistical validations, including linearity, error normality, absence of autocorrelation and homoscedasticity, which guaranteed their reliability. This hybrid approach proved effective in optimizing the predictive performance of PV systems under challenging conditions. Future work will explore the integration of these models with energy storage systems and smart control strategies to improve operational stability. In addition, the application of the hybrid model in extreme climates, such as desert or polar areas, will be investigated, as well as its extension through deep learning techniques to capture non-linear relationships and increase adaptability to abrupt climate variations.

PMID:40388424 | DOI:10.1371/journal.pone.0324047

PLoS One. 2025 May 19;20(5):e0317519. doi: 10.1371/journal.pone.0317519. eCollection 2025.

ABSTRACT

This study significantly contributes to the sphere of educational technology by deploying state-of-the-art machine learning and deep learning strategies for meaningful changes in education. The hybrid stacking approach did an excellent implementation using Decision Trees, Random Forest, and XGBoost as base learners with Gradient Boosting as a meta-learner, which managed to record an accuracy of 90%. That indeed puts into great perspective the huge potential it possesses for accuracy measures while predicting in educational setups. The CNN model, which predicted with an accuracy of 89%, showed quite impressive capability in sentiment analysis to acquire further insight into the emotional status of the students. RCNN, Random Forests, and Decision Trees contribute to the possibility of educational data complexity with valuable insight into the complex interrelationships within ML models and educational contexts. The application of the bagging XGBoost algorithm, which attained a high accuracy of 88%, further stamps its utility toward enhancement of academic performance through strong robust techniques of model aggregation. The dataset that was used in this study was sourced from Kaggle, with 1205 entries of 14 attributes concerning adaptability, sentiment, and academic performance; the reliability and richness of the analytical basis are high. The dataset allows rigorous modeling and validation to be done to ensure the findings are considered robust. This study has several implications for education and develops on the key dimensions: teacher effectiveness, educational leadership, and well-being of the students. From the obtained information about student adaptability and sentiment, the developed system helps educators to make modifications in instructional strategy more efficiently for a particular student to enhance effectiveness in teaching. All these aspects could provide critical insights for the educational leadership to devise data-driven strategies that would enhance the overall school-wide academic performance, as well as create a caring learning atmosphere. The integration of sentiment analysis within the structure of education brings an inclusive, responsive attitude toward ensuring students' well-being and, thus, a caring educational environment. The study is closely aligned with sustainable ICT in education objectives and offers a transformative approach to integrating AI-driven insights with practice in this field. By integrating notorious ML and DL methodologies with educational challenges, the research puts the basis for future innovations and technology in this area. Ultimately, it contributes to sustainable improvement in the educational system.

PMID:40388422 | DOI:10.1371/journal.pone.0317519

PLoS One. 2025 May 19;20(5):e0316043. doi: 10.1371/journal.pone.0316043. eCollection 2025.

ABSTRACT

The adoption of deep learning in ECG diagnosis is often hindered by the scarcity of large, well-labeled datasets in real-world scenarios, leading to the use of transfer learning to leverage features learned from larger datasets. Yet the prevailing assumption that transfer learning consistently outperforms training from scratch has never been systematically validated. In this study, we conduct the first extensive empirical study on the effectiveness of transfer learning in multi-label ECG classification, by investigating comparing the fine-tuning performance with that of training from scratch, covering a variety of ECG datasets and deep neural networks. Firstly, We confirm that fine-tuning is the preferable choice for small downstream datasets; however, it does not necessarily improve performance. Secondly, the improvement from fine-tuning declines when the downstream dataset grows. With a sufficiently large dataset, training from scratch can achieve comparable performance, albeit requiring a longer training time to catch up. Thirdly, fine-tuning can accelerate convergence, resulting in faster training process and lower computing cost. Finally, we find that transfer learning exhibits better compatibility with convolutional neural networks than with recurrent neural networks, which are the two most prevalent architectures for time-series ECG applications. Our results underscore the importance of transfer learning in ECG diagnosis, yet depending on the amount of available data, researchers may opt not to use it, considering the non-negligible cost associated with pre-training.

PMID:40388401 | DOI:10.1371/journal.pone.0316043

PLoS One. 2025 May 19;20(5):e0320426. doi: 10.1371/journal.pone.0320426. eCollection 2025.

ABSTRACT

Monitoring the remaining food in patients' trays is a routine activity in healthcare facilities as it provides valuable insights into the patients' dietary intake. However, estimating food leftovers through visual observation is time-consuming and biased. To tackle this issue, we have devised an efficient deep learning-based approach that promises to revolutionize how we estimate food leftovers. Our first step was creating the LeFoodSet dataset, a pioneering large-scale open dataset explicitly designed for estimating food leftovers. This dataset is unique in its ability to estimate leftover rates and types of food. To the best of our knowledge, this is the first comprehensive dataset for this type of analysis. The dataset comprises 524 image pairs representing 34 Indonesian food categories, each with images captured before and after consumption. Our prediction models employed a combined visual feature extraction and late fusion approach utilizing soft parameter sharing. Here, we used multi-task (MT) models that simultaneously predict leftovers and food types in training. In the experiments, we tested the single task (ST) model, the ST Model with Ground Truth (ST-GT), the MT model, and the MT model with Inter-task Connection (MT-IC). Our AI-based models, particularly the MT and MT-IC models, have shown promising results, outperforming human observation in predicting leftover food. These findings show the best with the ResNet101 model, where the Mean Average Error (MAE) of leftover task and food classification accuracy task is 0.0801 and 90.44% in the MT Model and 0.0817 and 92.56% in the MT-IC Model, respectively. It is proved that the proposed solution has a bright future for AI-based approaches in medical and nursing applications.

PMID:40388400 | DOI:10.1371/journal.pone.0320426

Bioinformatics. 2025 May 19:btaf165. doi: 10.1093/bioinformatics/btaf165. Online ahead of print.

ABSTRACT

MOTIVATION: Deep learning has deeply influenced protein science, enabling breakthroughs in predicting protein properties, higher-order structures, and molecular interactions.

RESULTS: This paper introduces DeepProtein, a comprehensive and user-friendly deep learning library tailored for protein-related tasks. It enables researchers to seamlessly address protein data with cutting-edge deep learning models. To assess model performance, we establish a benchmark that evaluates different deep learning architectures across multiple protein-related tasks, including protein function prediction, subcellular localization prediction, protein-protein interaction prediction, and protein structure prediction. Furthermore, we introduce DeepProt-T5, a series of fine-tuned Prot-T5-based models that achieve state-of-the-art performance on four benchmark tasks, while demonstrating competitive results on six of others. Comprehensive documentation and tutorials are available which could ensure accessibility and support reproducibility.

AVAILABILITY AND IMPLEMENTATION: Built upon the widely used drug discovery library DeepPurpose, DeepProtein is publicly available at https://github.com/jiaqingxie/DeepProtein.

SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

PMID:40388205 | DOI:10.1093/bioinformatics/btaf165

Interdiscip Cardiovasc Thorac Surg. 2025 May 19:ivaf101. doi: 10.1093/icvts/ivaf101. Online ahead of print.

ABSTRACT

OBJECTIVES: This study aimed to develop an automated method for pulmonary artery and vein segmentation in both left and right lungs from computed tomography (CT) images using artificial intelligence (AI). The segmentations were evaluated using PulmoSR software, which provides 3D visualizations of patient-specific anatomy, potentially enhancing a surgeon's understanding of the lung structure.

METHODS: A dataset of 125 CT scans from lung segmentectomy patients at Erasmus MC was used. Manual annotations for pulmonary arteries and veins were created with 3D Slicer. nnU-Net models were trained for both lungs, assessed using Dice score, sensitivity, and specificity. Intraoperative recordings demonstrated clinical applicability. A paired t-test evaluated statistical significance of the differences between automatic and manual segmentations.

RESULTS: The nnU-Net model, trained at full 3D resolution, achieved a mean Dice score between 0.91 and 0.92. The mean sensitivity and specificity were: left artery: 0.86 and 0.99, right artery: 0.84 and 0.99, left vein: 0.85 and 0.99, right vein: 0.85 and 0.99. The automatic method reduced segmentation time from ∼1.5 hours to under 5 min. Five cases were evaluated to demonstrate how the segmentations support lung segmentectomy procedures. P-values for Dice scores were all below 0.01, indicating statistical significance.

CONCLUSIONS: The nnU-Net models successfully performed automatic segmentation of pulmonary arteries and veins in both lungs. When integrated with visualization tools, these automatic segmentations can enhance preoperative and intraoperative planning by providing detailed 3D views of patients anatomy.

PMID:40388152 | DOI:10.1093/icvts/ivaf101

ACS Nano. 2025 May 19. doi: 10.1021/acsnano.5c04340. Online ahead of print.

ABSTRACT

Conventional point-of-care testing (POCT) has limitations in sensitivity with high risks of missed detection or false positive, which restrains its applications for routine outpatient care analysis and early clinical diagnosis. By merits of the cutting-edge quantum precision metrology, this study devised a mini quantum sensor via magnetofluorescent qubit tagging and tunning on core-shelled fluorescent nanodiamond FND@SiO2. Comprehensive characterizations confirmed the formation of FND biolabels, while spectroscopies secured no degradation in spin-state transition after surface modification. A methodical parametrization was deliberated and decided, accomplishing a wide-field modulation depth ≥15% in ∼ zero field, which laid foundation for supersensitive sensing at single-FND resolution. Using viral nucleocapsid protein as a model marker, an ultralow limit of detection (LOD) was obtained by lock-in analysis, outperforming conventional colorimetry and immunofluorescence by > 2000 fold. Multianalyte and affinity assays were also enabled on this platform. Further by resort to artificial-intelligence (AI) augmentation in the Unet-ConvLSTM-Attention architecture, authentic qubit dots were identified by pixelwise survey through pristine qubit queues. Such processing not just improved pronouncedly the probing precision but also achieved deterministic detections down to a single protein in human saliva with an ultimate LOD as much as 7800-times lower than that of colloidal Au approach, which competed with the RT-qPCR threshold and the certified critical value of SIMOA, the gold standard. Hence, by AI-aided digitization on optic qubits, this REASSURED-compliant contraption may promise a next-generation POCT solution with unparalleled sensitivity, speed, and cost-effectiveness, which in whole confers a conclusive proof of the prowess of the burgeoning quantum metrics in biosensing.

PMID:40388114 | DOI:10.1021/acsnano.5c04340