Int J Comput Assist Radiol Surg. 2025 May 3. doi: 10.1007/s11548-025-03356-7. Online ahead of print.

ABSTRACT

PURPOSE: Quantitative magnetic resonance imaging (qMRI) enables imaging of physical parameters related to the nuclear spin of protons in tissue, and is poised to revolutionize clinical research. However, improving the accuracy and clinical relevance of qMRI is essential for its practical implementation. This requires significantly reducing the currently lengthy acquisition times to enable clinical examinations and provide an environment where clinical accuracy and reliability can be verified. Deep learning (DL) has shown promise in significantly reducing imaging time and improving image quality in recent years. This study introduces a novel approach, quantitative deep cascade of convolutional network (qDC-CNN), as a framework for accelerated quantitative parameter mapping, offering a potential solution to this challenge. This work aims to verify that the proposed model outperforms the competing methods.

METHODS: The proposed qDC-CNN is an integrated deep-learning framework combining an unrolled image reconstruction network and a fully connected neural network for parameter estimation. Training and testing utilized simulated multi-slice multi-echo (MSME) datasets generated from the BrainWeb database. The reconstruction error with ground truth was evaluated using normalized root mean squared error (NRMSE) and compared with conventional DL-based methods. Two validation experiments were performed: (Experiment 1) assessment of acceleration factor (AF) dependency (AF = 5, 10, 20) with fixed 16 echoes, and (Experiment 2) evaluation of the impact of reducing contrast images (16, 8, 4 images).

RESULTS: In most cases, the NRMSE values of S0 and T2 estimated from the proposed qDC-CNN were within 10%. In particular, the NRMSE values of T2 were much smaller than those of the conventional methods.

CONCLUSIONS: The proposed model had significantly smaller reconstruction errors than the conventional models. The proposed method can be applied to other qMRI sequences and has the flexibility to replace the image reconstruction module to improve performance.

PMID:40317423 | DOI:10.1007/s11548-025-03356-7

ACS Nano. 2025 May 3. doi: 10.1021/acsnano.4c15032. Online ahead of print.

ABSTRACT

The search for and design of superconductors with both Kagome lattice and hardness is a challenging and frontier research topic. This work utilizes structure predictions to discover the Kagome lattice in NaSi3_P6/mmm phase of NaxSiy (x, y = 1-3). For a comprehensive understanding of XY3_P6/mmm, other atoms such as X = Li, Na, Cs and Y = B, Si, Ge are included. Superconducting critical temperatures (Tc) of XY3 compounds are calculated between 0 and 20 GPa and found to be 30.54 K for CsB3 at 0 GPa, indicating that electron-phonon coupling, phonon softening, linewidths, and electron density at the Fermi level all have significant effects on Tc. The bonding type of B, Si, and Ge atoms in the Kagome lattice also determines the boundaries of its hard properties and superconductivity. Moreover, the melting temperature of NaSi3_P6/mmm is determined to be 608 K at 0 GPa and P-T phase diagram at pressures of 0-15 GPa using deep learning molecular dynamics simulations. Our findings provide a multitude of excellent properties in the XY3 compounds, including Kagome lattice, high hardness, and superconductors, which will provide essential physical insights and theoretical guidance for the experimental exploration of the hard superconductors.

PMID:40317254 | DOI:10.1021/acsnano.4c15032

J Cancer Res Ther. 2025 May 1;21(2):362-370. doi: 10.4103/jcrt.jcrt_1910_24. Epub 2025 May 2.

ABSTRACT

BACKGROUND: To develop a deep learning radiomics (DLR) model based on contrast-enhanced computed tomography (CECT) to assess the rat sarcoma (RAS) oncogene status and predict targeted therapy response in colorectal cancer liver metastases (CRLM).

METHODS: This multicenter retrospective study comprised 185 CRLM patients who were categorized into three cohorts: training (n = 88), internal test (n = 39), and external test (n = 58). A total of 1126 radiomic features and 2589 DL signatures were extracted from each region of interest in the CECT. Fourteen significant radiomic features associated with RAS mutation were selected. Subsequently, various models (DL-arterial phase (AP), DL-venous phase (VP), AP+VP-DL, radiomics, and DL-R) were developed and validated. The model performance was compared using the area under the receiver operating characteristic (AUROC) curves and the DeLong test. The predictive usefulness of the DL score for progression-free survival and overall survival (OS) was determined.

RESULTS: The AP+VP-DL model achieved the highest AUC (0.98), outperforming the radiomics (0.90), DL-AP (0.93), DL-VP (0.87), and DL-R (0.97) models. Significant associations were observed between OS and the carcinoembryonic antigen (CEA), disease control rate (DCR), and DL scores, leading to the development of a DL nomogram. A high-risk RAS mutation status correlated with significantly lower 1-year (88% vs. 96%), 3-year (12% vs. 35%), and 5-year (0% vs. 15%) cumulative survival rates compared to a low-risk status (P = 0.03).

CONCLUSIONS: The DL model demonstrated satisfactory predictive performance, aiding clinicians in noninvasively predicting the RAS gene status for informed treatment decisions.

PMID:40317140 | DOI:10.4103/jcrt.jcrt_1910_24

Ultrasound Med Biol. 2025 May 1:S0301-5629(25)00094-8. doi: 10.1016/j.ultrasmedbio.2025.03.015. Online ahead of print.

ABSTRACT

Cardiovascular diseases (CVDs) remain a leading cause of mortality worldwide, underscoring the need for advancements in diagnostic methodologies to improve early detection and treatment outcomes. This systematic review examines the integration of advanced deep learning (DL) techniques in echocardiography for detecting cardiovascular abnormalities, adhering to PRISMA 2020 guidelines. Through a comprehensive search across databases like IEEE Xplore, PubMed, and Web of Science, 29 studies were identified and analyzed, focusing on deep convolutional neural networks (DCNNs) and their role in enhancing the diagnostic precision of echocardiographic assessments. The findings highlight DL's capability to improve the accuracy and reproducibility of detecting and classifying echocardiographic data, particularly in measuring left ventricular function and identifying wall motion abnormalities. Despite these advancements, challenges such as data diversity, image quality, and the computational demands of DL models hinder their broader clinical adoption. In conclusion, DL offers significant potential to enhance the diagnostic capabilities of echocardiography. However, successful clinical implementation requires addressing issues related to data quality, computational demands, and system integration.

PMID:40316488 | DOI:10.1016/j.ultrasmedbio.2025.03.015

BMJ Open. 2025 May 2;15(5):e094764. doi: 10.1136/bmjopen-2024-094764.

ABSTRACT

INTRODUCTION: Updated primary prevention strategies are needed for post-infarction sudden cardiac death (SCD) based on implantable cardioverter-defibrillator (ICD). Current recommendations, based on left ventricular systolic function and functional class, may be obsolete because they are derived from ancient studies that do not incorporate the potential benefit of either current comprehensive treatment of ischaemic heart disease or modern device programming. Among patients with post-infarction left ventricular dysfunction, modern implantable cardiac monitoring devices (ICM) allow a unique opportunity to determine in real-time the burden of non-sustained ventricular tachycardias and their relationship to the subsequent occurrence of sustained or symptomatic events.

METHODS AND ANALYSIS: Approximately 200 patients with left ventricular ejection fraction (LVEF) equal to or less than 40% after acute myocardial infarction will be included in the study. They will be implanted with a Confirm RX, an ICM with real-time remote connection via a smartphone. At 6 months, LVEF and functional status will be re-evaluated and cardiac morpho-functional characterisation will be performed by MRI. At this time, and following current European guidelines, patients with an indication will receive an ICD; the others will continue to be monitored using an ICM for a minimum of 2 years. Patients are expected to be followed up for 4 years after the index event. More than 20 000 remote transmissions are expected to be analysed. The study will focus on the relationship between the detection of non-sustained ventricular tachycardias by ICMs (defined as at least 8 R-R intervals at 160 beats per minute) and the subsequent occurrence of symptomatic arrhythmic events. An advanced statistical analysis will be performed using machine and deep learning techniques to determine the clinical variables, those that are derived from monitoring and imaging tests and related to mid-term prognosis.

ETHICS AND DISSEMINATION: The study was approved by the Ethical Committee of the University Hospital of Salamanca (protocol number PI 2019 03 246) on 30 April 2020. Each patient will be informed about the study in both oral and written form by a physician and will be included in the study after written consent is obtained.For the first time, a study will provide real-time information on the arrhythmic burden of patients with post-infarction ventricular dysfunction and its prognostic implications in the medium term. Several publications in scientific journals are planned.

TRIAL REGISTRATION NUMBER: NCT04765943.

PMID:40316360 | DOI:10.1136/bmjopen-2024-094764

Neuroimage. 2025 Apr 30:121246. doi: 10.1016/j.neuroimage.2025.121246. Online ahead of print.

ABSTRACT

OBJECTIVES: Multiplexed Positron Emission Tomography (PET) imaging allows simultaneous acquisition of multiple radiotracer signals, thus enhancing diagnostic capabilities, reducing scan times, and improving patient comfort. Traditional methods often require significant delays between tracer injections, leading to physiological changes and noise interference. Recent advancements, including multi-tracer compartment modeling and machine learning, provide promising solutions. This study explores the deep learning (DL)-based single-session triple-tracer brain PET imaging protocol, aiming at simplifying multi-tracer PET imaging, while reducing radiation exposure.

METHODS: The study uses the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset, which includes cognitively normal (CN) patients, as well as patients with mild cognitive impairment (MCI) and dementia. The dataset also includes PET scans acquired with amyloid (18F-florbetaben [FBB] or 18F-florbetapir [FBP]), 18F-Fluorodeoxyglucose (FDG), and tau 18F-flortaucipir (FTP). To mimic the effect of simultaneous acquisition of multiple PET tracers, we generated synthetic dual- and triple-tracer images by summing FBP/FBB, FTP, and FDG scans. A DL model based on Swin Transformer architecture was developed to separate these signals, using five-fold cross-validation and mean squared error (MSE) loss. The synthetic PET images were evaluated using established image quality metrics, including MSE, structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR). In addition, clinical evaluation was conducted by two nuclear medicine specialists to assess the amyloid and tau status in the synthetic and reference images.

RESULTS: The proposed DL model effectively synthesized realistic FBB/FBP and FDG images from dual- and triple-tracer PET images. Although the proposed DL model's performance in generating FTP images was less successful, it remains promising. The clinical evaluation revealed that the amyloid status estimated from the synthetic images led to a sensitivity of 92% and specificity of 86% for FBB, while it showed a sensitivity of 93% and specificity of 67% for tau status using FBP extracted from the triple-tracer images. The calculated quantitative metrics showed that the mean error for synthetic amyloid images (FBB: 0.03 SUV, FBP: 0.00 SUV) was higher than FDG for FBB (0.02 SUV) but lower than FDG for FBP (-0.01 SUV), and comparable to FTP (FBB: 0.03 SUV, FBP: 0.00 SUV). Voxel-wise correlation analysis demonstrated strong correlation between synthetic and reference images, particularly for amyloid images (FBB: y = 0.98x + 0.00, R² = 0.85; FBP: y = 1.11x + 0.04, R² = 0.73), while FTP (FBB: y = 0.87x + 0.14, R² = 0.51; FBP: y = 0.98x + 0.09, R² = 0.59) and FDG images (FBB: y = 1.01x + 0.18, R² = 0.85; FBP: y = 0.96x + 1.37, R² = 0.77) showed moderate correlations.

CONCLUSION: Our study demonstrates that the suggested DL model can separate the signals belonging to three different radiotracers from simultaneous triple-tracer PET scans. This method may make multiplex scanning feasible in the clinic, hence reducing the scanning time, radiation hazard and improving patient comfort.

PMID:40316225 | DOI:10.1016/j.neuroimage.2025.121246

J Neurosci Methods. 2025 Apr 30:110463. doi: 10.1016/j.jneumeth.2025.110463. Online ahead of print.

ABSTRACT

BACKGROUND: Taste perception is the process by which the gustatory system detects and interprets chemical stimuli from food and beverages, involving activation of taste receptors on the tongue. Analyzing taste perception is essential for understanding human sensory responses and diagnosing taste-related disorders.

NEW METHOD: This research focuses on developing a deep learning framework to effectively recognize basic taste stimuli from EEG signals. Initially, the recorded EEG signals undergo preprocessing to remove noise and artifacts. The CEEMDAN (complete ensemble empirical mode decomposition with adaptive noise) method is then applied to decompose the EEG signals into various frequency rhythms, referred to as intrinsic mode functions (IMFs). From the chosen IMFs, six distinct entropy features-sample, bubble, approximate, dispersion, slope, and permutation entropy-are extracted for further analysis. A novel deep learning model, TasteNet, is then developed, integrating a convolutional neural network (CNN) module, a multi-head attention module, and the Att-BiPLSTM (Attention-Bidirectional Potent Long Short-Term Memory) network.

RESULTS: The proposed architecture classifies the input data into six categories: no taste, sweet, sour, bitter, umami, and salty, achieving a remarkable accuracy of 97.52 ±0.48%.

COMPARISON WITH EXISTING METHODS: TasteNet outperforms existing taste perception classification methods, as demonstrated through extensive experiments.

CONCLUSION: This study presents TasteNet, a robust framework for precise taste perception recognition using EEG signals. Using CEEMDAN for effective signal decomposition and extracting key entropy features, the model captures intricate patterns in taste stimuli. The incorporation of multi-head attention module and the Att-BiPLSTM network further enhances the model's ability to identify various taste sensations accurately.

PMID:40315923 | DOI:10.1016/j.jneumeth.2025.110463

Comput Biol Med. 2025 May 1;192(Pt A):110234. doi: 10.1016/j.compbiomed.2025.110234. Online ahead of print.

ABSTRACT

The widespread use of immersive technologies such as Virtual Reality, Mixed Reality, and Augmented Reality has led to the continuous collection and streaming of vast amounts of sensitive biometric data. Among the biometric signals collected, ECG (electrocardiogram) stands out given its critical role in healthcare, particularly for the diagnosis and management of cardiovascular diseases. Numerous studies have demonstrated that ECG contains traits to distinctively identify a person. As a result, the need for anonymization methods is becoming increasingly crucial to protect personal privacy while ensuring the integrity of health data for effective clinical utility. Although many anonymization methods have been proposed in the literature, there has been limited exploration into their ability to preserve data integrity while complying with stringent data protection regulations. More specifically, the utility of anonymized signal and the privacy level achieved often present a trade-off that has not been thoroughly addressed. This paper analyzes the trade-off between balancing privacy protection with the preservation of health data integrity in ECG signals focusing on memory-efficient anonymization techniques that are suitable for real-time or streaming applications and do not require heavy memory computation. Moreover, we introduce an analytical framework to evaluate the privacy preservation methods alongside health integrity, incorporating state-of-the-art disease and person identifiers. We also propose a novel metric that assists users in selecting an anonymization method based on their desired trade-off between health insights and privacy protection. The experimental results demonstrate the impact of the de-identification techniques on critical downstream tasks, such as Arrhythmia detection and Myocardial Infarction detection along with identification performance, while statistical analysis reveals the biometric nature of ECG signals. The findings highlight the limitations of using such anonymization methods and models, emphasizing the need for approaches that maintain the clinical relevance of ECG data in real-time and streaming applications, particularly in memory-constrained environments.

PMID:40315720 | DOI:10.1016/j.compbiomed.2025.110234

Comput Biol Med. 2025 May 1;192(Pt A):110273. doi: 10.1016/j.compbiomed.2025.110273. Online ahead of print.

ABSTRACT

Plankton are microorganisms that play an important role in marine food webs as primary producers in the trophic web. Traditional plankton identification methods using manual microscopy and sampling are time-consuming, labor-intensive, and prone to errors. Deep learning has improved the automation of plankton identification, but it remains challenging to achieve high accuracy and efficiency in computation with limited labeled data. In this paper, we proposed an improved plankton classification model that is more accurate and interpretable. We train two models, InceptionResNetV2 (transfer learning) and DeepPlanktonNet (from scratch), on the WHOI dataset. We utilize feature fusion to supplement feature representation, merging the outputs of both models. Feature selection is achieved through the Whale Optimization Algorithm (WOA), eliminating redundancy and making it more computationally efficient. Additionally, we also employ Local Interpretable Model-agnostic Explanations (LIME) to make the model more interpretable and gain insights into how the model makes decisions. Additionally, feature selection using WOA reduces feature space and has less inference and computational cost. Our method achieves a classification accuracy of 98.79 %, which is better than previous state-of-the-art methods. For robustness testing, we train nine machine learning classifiers on the optimized features. By significantly improving classification accuracy and speed, our method enables large-scale ecological surveys, water quality monitoring, and biodiversity studies. These advances allow researchers to and environmental scientists to automate plankton classification more reliably, supporting marine conservation and resource management.

PMID:40315719 | DOI:10.1016/j.compbiomed.2025.110273

BMC Infect Dis. 2025 May 2;25(1):647. doi: 10.1186/s12879-025-10916-4.

ABSTRACT

INTRODUCTION: Pneumonia is the leading cause of child morbidity and mortality and accounts for 5.6 million under-five child deaths. Pneumonia has a significant impact on the quality of life, the country's economy, and the survival of children. Therefore, this study aimed to develop data-driven predictive model using machine learning algorithms to predict pneumonia and stratify the determinant factors among children aged 6-23 months in Ethiopia.

METHODS: A total of 2035 samples of children were used from the 2016 Ethiopian Demographic and Health Survey dataset. Jupyter Notebook from Anaconda Navigators was used for data management and analysis. Important libraries such as Pandas, Seaborn, and Numpy were imported from Python. The data was pre-processed into a training and testing dataset with a 4:1 ratio, and tenfold cross-validation was used to reduce bias and enhance the models' performance. Six machine learning algorithms were used for model building and comparison, and confusion matrix elements were used to evaluate the performance of each algorithm. Principal component analysis and heatmap function were used for correlation detection between features. Feature importance score was used to identify and stratify the most important predictors of pneumonia.

RESULTS: From 2035 total samples, 16.6%, 20.1%, and 24.2% of children had short rapid breath, fever, and cough respectively. The overall magnitude of pneumonia among children aged 6-23 months was 31.3% based on the 2016 EDHS report. A random forest algorithm is the relatively best performance model to predict pneumonia and stratify its determinates with 91.3% accuracy. The health facility visits, child sex, initiation of breastfeeding, birth interval, birth weight, husbands' education, women's age, and region, are the top eight important predictors of pneumonia among children with important scores of more than 5% to 20% respectively.

CONCLUSIONS: Random forest is the best model to predict pneumonia and stratify its determinant factors. The implications of this study are profound for advanced research methodology, tailored to promote effective health interventions such as lifestyle modification and behavioral intervention, based on individuals' unique features, specifically for stakeholders to take proactive childcare interventions. The study would serve as pioneering evidence for future research, and researchers are recommended to use deep learning algorithms to enhance prediction accuracy.

PMID:40316929 | DOI:10.1186/s12879-025-10916-4

Mol Divers. 2025 May 2. doi: 10.1007/s11030-025-11206-6. Online ahead of print.

ABSTRACT

Human metapneumovirus (hMPV) primarily causes respiratory tract infections in young children and older adults. According to the 2024 Human Pneumonia Etiology Research for Child Health (PERCH) study, hMPV is the second leading common cause of pneumonia in children under five in Asia and Africa. The virus encodes nine proteins, including the essential Fusion (F) and G glycoproteins, which facilitate entry to the host cells. Currently, there are no approved vaccines or antiviral treatments for hMPV; supportive care is the primary way it is managed. Hence, this study focuses on the F protein as a therapeutic target to find a repurposable drug to fight hMPV. Refolding of the F protein and its binding to heparan sulfate enable hMPV infection. Heparin sulfate is important for hMPV binding, and we have found that cangrelor and AVN 944 can prevent the fusion of membranes. We developed a deep learning-based pharmacophore to identify potential drugs targeting hMPV, from which we could narrowed a list of 2400 FDA-approved drugs and 255 antiviral drugs to 792 and 72 drugs, respectively. We then conducted quantitative validation using the ROC curve. Further virtual screening of the drugs was performed, leading us to select the one with the highest docking score. The validation of the deep learning prediction in virtual screening Pearson correlation was done. Further, the MD simulation of these drugs confirmed that the protein-drug complex stability remained in dynamic condition. Further, the stability of protein-drug complexes than unbound protein was confirmed by Free Energy Landscape and Dynamic Cross Correlation Matrices. Further in vitro and in vivo experiments need to determine the efficacy of the identified candidates.

PMID:40316857 | DOI:10.1007/s11030-025-11206-6

Sci Rep. 2025 May 2;15(1):15432. doi: 10.1038/s41598-025-99676-4.

ABSTRACT

Secure medical data sharing and access control play a prominent role. However, it is still unclear how to provide a security architecture that can guarantee the privacy and safety of sensitive medical data. Existing methods are application-specific and fail to take into account the complex security needs of healthcare applications. Moreover, the healthcare sector needs dynamic permission enforcement, extensible context-aware access control, flexible, and on-demand authentication. Therefore, this research proposes an access control mechanism and an effective attack detection model. The proposed authenticate access control mechanism (PA2C) safeguards data integrity as well as the security and dependability of EHR data sharing are improved by the use of smart contracts, encryption, and secure key management. On the other hand, the proposed intelligent voyage optimization algorithm-based Radial basis neural network (IntVO-RBNN) effectively detects the attacks in the network. Specifically, the Intelligent Voyage Optimization algorithm effectively tunes the model hyperparameters and the deployment of hybrid features contributes to the proposed model to detect attack patterns effectively. The comparative results showed that the suggested access control strategy performed better than the current methods in terms of minimal responsiveness of 100.18 s and less information loss of 4.49% for 100 blocks. Likewise, the proposed IntVO-RBNN attack detection model performs better with 95.26% recall, 97.84% precision, and 94.02% accuracy.

PMID:40316724 | DOI:10.1038/s41598-025-99676-4

Sci Rep. 2025 May 2;15(1):15333. doi: 10.1038/s41598-025-99701-6.

ABSTRACT

Patients with cardioembolic stroke often undergo CT of the left atrial appendage (LAA), for example, to determine whether thrombi are present in the LAA. To guide the imaging process, technologists first perform a localizer scan, which is a preliminary image used to identify the region of interest. However, the lack of well-defined landmarks makes accurate delimitation of the LAA in localizers difficult and often requires whole-heart scans, increasing radiation exposure and cancer risk. This study aims to automate LAA delimitation in CT localizers using deep learning. Four commonly used deep networks (VariFocalNet, Cascade-R-CNN, Task-aligned One-stage Object Detection Network, YOLO v11) were trained to predict the LAA boundaries on a cohort of 1253 localizers, collected retrospectively from a single center. The best-performing network in terms of delimitation accuracy was then evaluated on an internal test cohort of 368 patients, and on an external test cohort of 309 patients. The VariFocalNet performed best, achieving LAA delimitations with high accuracy (97.8% and 96.8%; Dice coefficients: 90.4% and 90.0%) and near-perfect clinical utility (99.8% and 99.3%). Compared to whole-heart scanning, the network-based delimitation reduced the radiation exposure by more than 50% (5.33 ± 6.42 mSv vs. 11.35 ± 8.17 mSv in the internal cohort, 4.39 ± 4.23 mSv vs. 10.09 ± 8.0 mSv in the external cohort). This study demonstrates that a deep learning network can accurately delimit the LAA in the localizer, leading to more accurate CT scans of the LAA, thereby significantly reducing radiation exposure to the patient compared to whole-heart scanning.

PMID:40316718 | DOI:10.1038/s41598-025-99701-6

Sci Rep. 2025 May 2;15(1):15428. doi: 10.1038/s41598-025-99436-4.

ABSTRACT

The research introduced a new method for land-use classification by merging deep convolutional neural networks with a modified variant of a metaheuristic optimization technique. The methodology involved utilizing the VGG-19 model for feature extraction, dimensionality reduction, and a stacked autoencoder optimized with a boosted version of the Big Bang Crunch Theory. Through testing on the Aerial Image Dataset the and UC Merced Land Use Dataset and comparing it with other published works, the approach showed higher classification accuracy compared to current state-of-the-art methods. The study revealed that incorporating boosted Big-Bang Crunch significantly enhances the performance of stacked autoencoder in land-use classification tasks. Moreover, comparisons with other techniques, including convolutional neural networks, Cascaded Residual Dilated Networks, hierarchical convolutional recurrent neural networks, Fusion Region Proposal Networks, and multi-level context-guided classification techniques using Object-Based Convolutional Neural Networks, emphasized the benefits of using Convolutional Neural Network models over traditional methods. The proposed model achieved an accuracy of 92.49% on the AID dataset and 95.93% on the UC Merced dataset, with precision scores of 98.64% and 98.93%, respectively. These results emphasize the importance of integrating deep learning architectures with sophisticated optimization techniques, contributing to enhanced land-use classification accuracy.

PMID:40316651 | DOI:10.1038/s41598-025-99436-4

Sci Rep. 2025 May 2;15(1):15345. doi: 10.1038/s41598-025-98389-y.

ABSTRACT

Electroencephalography (EEG) is one of the most used techniques to perform diagnosis of epilepsy. However, manual annotation of seizures in EEG data is a major time-consuming step in the analysis process of EEGs. Different machine learning models have been developed to perform automated detection of seizures from EEGs. However, a large gap is observed between initial accuracies and those observed in clinical practice. In this work, we reproduced and assessed the accuracy of a large number of models, including deep learning networks, for detection of seizures from EEGs. Benchmarking included three different datasets for training and initial testing, and a manually annotated EEG from a local patient for further testing. Random forest and a convolutional neural network achieved the best results on public data, but a large reduction of accuracy was observed testing with the local data, especially for the neural network. We expect that the retrained models and the data available in this work will contribute to the integration of machine learning techniques as tools to improve the accuracy of diagnosis in clinical settings.

PMID:40316648 | DOI:10.1038/s41598-025-98389-y

Sci Rep. 2025 May 2;15(1):15387. doi: 10.1038/s41598-025-98264-w.

ABSTRACT

Mental disorders represent a critical global health challenge that affects millions around the world and significantly disrupts daily life. Early and accurate detection is paramount for timely intervention, which can lead to improved treatment outcomes. Electroencephalography (EEG) provides the non-invasive means for observing brain activity, making it a useful tool for detecting potential mental disorders. Recently, deep learning techniques have gained prominence for their ability to analyze complex datasets, such as electroencephalography recordings. In this study, we introduce a novel deep-learning architecture for the classification of mental disorders such as post-traumatic stress disorder, depression, or anxiety, using electroencephalography data. Our proposed model, the multichannel convolutional transformer, integrates the strengths of both convolutional neural networks and transformers. Before feeding the model as low-level features, the input is pre-processed using a common spatial pattern filter, a signal space projection filter, and a wavelet denoising filter. Then the EEG signals are transformed using continuous wavelet transform to obtain a time-frequency representation. The convolutional layers tokenize the input signals transformed by our pre-processing pipeline, while the Transformer encoder effectively captures long-range temporal dependencies across sequences. This architecture is specifically tailored to process EEG data that has been preprocessed using continuous wavelet transform, a technique that provides a time-frequency representation, thereby enhancing the extraction of relevant features for classification. We evaluated the performance of our proposed model on three datasets: the EEG Psychiatric Dataset, the MODMA dataset, and the EEG and Psychological Assessment dataset. Our model achieved classification accuracies of 87.40% on the EEG and Psychological Assessment dataset, 89.84% on the MODMA dataset, and 92.28% on the EEG Psychiatric dataset. Our approach outperforms every concurrent approaches on the datasets we used, without showing any sign of over-fitting. These results underscore the potential of our proposed architecture in delivering accurate and reliable mental disorder detection through EEG analysis, paving the way for advancements in early diagnosis and treatment strategies.

PMID:40316629 | DOI:10.1038/s41598-025-98264-w

Sci Rep. 2025 May 2;15(1):15404. doi: 10.1038/s41598-025-94239-z.

ABSTRACT

This paper provides an in-depth analysis and performance evaluation of four Solar Radiance (SR) prediction models. The prediction is ensured for a period ranging from a few hours to several days of the year. These models are derived from four machine learning methods, namely the Feed-forward Back Propagation (FFBP) method, Convolutional Feed-forward Back Propagation (CFBP) method, Support Vector Regression (SVR), and the hybrid deep learning (DL) method, which combines Convolutional Neural Networks and Long Short-Term Memory networks. This combination results in the CNN-LSTM model. Additionally, statistical indicators use Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and Normalized Root Mean Squared Error (nRMSE). Each indicator compares the predicted output by each model above and the actual output, pre-recorded in the experimental trial. The experimental results consistently show the power of the CNN-LSTM model compared to the remaining models in terms of accuracy and reliability. This is due to its lower error rate and higher detection coefficient (R2 = 0.99925).

PMID:40316622 | DOI:10.1038/s41598-025-94239-z

Sci Rep. 2025 May 2;15(1):15411. doi: 10.1038/s41598-025-96588-1.

ABSTRACT

This study explores the motivations behind integrating TinyML-based voice assistants into daily life, focusing on enhancing their user interface (UI) and functionality to improve user experience. This research discusses real-world applications like smart home automation, visually impaired assistive technologies, and healthcare monitoring. This review acknowledges various problems and helps us understand why TinyML exerts such significant implications in numerous domains. Researchers derive solutions from this study on how voice assistants integrated with TinyML can effectively analyze and adjust to user behaviour patterns in real-world scenarios, thereby enabling the delivery of dynamic and responsive content to enhance user engagement. The article also focused on limitations while implementing TinyML. Researchers will understand the detailed issues that are unavailable in most papers. This work explores features that can be embedded in voice assistants, like smart home automation, smart watches, smart glasses for visually impaired people, etc., using TinyML. A comparative review of current methods identifies areas of research gaps such as deployment difficulties, noise interference, and model efficiency on low-resource devices. From this study, researchers can directly identify the research gap with minimal effort, which may motivate them to focus more on solving the open problems due to optimize the problem identification time.

PMID:40316605 | DOI:10.1038/s41598-025-96588-1

Sci Rep. 2025 May 2;15(1):15412. doi: 10.1038/s41598-025-00331-9.

ABSTRACT

Smart weed-crop discrimination is crucial for modern precision weed management. In this study, we aimed to develop a robust system for site-specific weed control in saffron fields by utilizing color images and a deep learning approach to distinguish saffron from four common weeds: flixweed, hoary cress, mouse barley, and wild garlic. A total of 504 images were taken in natural and unstructured field settings. Eight state-of-the-art deep learning networks - VGG19, ResNet152, Xception, InceptionResNetV2, EfficientNetB0, EfficientNetB1, EfficientNetV2B0, and EfficientNetV2B1 were evaluated as potential base networks. These networks underwent pre-training on ImageNet using transfer learning, followed by fine-tuning and improvement with additional layers to optimize performance on our dataset. The improved EfficientNetB0 model stood out as the top performer among the eight models, achieving an accuracy rate of 94.06% and a loss value of 0.513 on the test dataset. This proposed model excelled in accurately classifying plant categories, obtaining f1-scores ranging from 82 to 100%. We scrutinized fifteen scenarios of weed presence in saffron fields, focusing on various weed types, to propose efficient management tactics using the model. These discoveries lay the groundwork for precise saffron weed management strategies that reduce herbicide use, environmental impact, and boost yield and quality.

PMID:40316572 | DOI:10.1038/s41598-025-00331-9

Brief Bioinform. 2025 May 1;26(3):bbaf160. doi: 10.1093/bib/bbaf160.

ABSTRACT

Single-cell sequencing has advanced our understanding of cellular heterogeneity and disease pathology, offering insights into cellular behavior and immune mechanisms. However, extracting meaningful phenotype-related features is challenging due to noise, batch effects, and irrelevant biological signals. To address this, we introduce Deep scSTAR (DscSTAR), a deep learning-based tool designed to enhance phenotype-associated features. DscSTAR identified HSP+ FKBP4+ T cells in CD8+ T cells, which linked to immune dysfunction and resistance to immune checkpoint blockade in non-small cell lung cancer. It has also enhanced spatial transcriptomics analysis of renal cell carcinoma, revealing interactions between cancer cells, CD8+ T cells, and tumor-associated macrophages that may promote immune suppression and affect outcomes. In hepatocellular carcinoma, it highlighted the role of S100A12+ neutrophils and cancer-associated fibroblasts in forming tumor immune barriers and potentially contributing to immunotherapy resistance. These findings demonstrate DscSTAR's capacity to model and extract phenotype-specific information, advancing our understanding of disease mechanisms and therapy resistance.

PMID:40315434 | DOI:10.1093/bib/bbaf160