Sci Rep. 2025 Apr 26;15(1):14652. doi: 10.1038/s41598-025-99451-5.

ABSTRACT

Neuroblastoma presents a wide variety of clinical phenotypes, demonstrating different levels of benignity and malignancy among its subtypes. Early diagnosis is essential for effective patient management. Computed tomography (CT) serves as a significant diagnostic tool for neuroblastoma, utilizing machine vision imaging, which offers advantages over traditional X-ray and ultrasound imaging modalities. However, the high degree of similarity among neuroblastoma subtypes complicates the diagnostic process. In response to these challenges, this study presents a modified version of the You Only Look Once (YOLO) algorithm, called YOLOv8-IE. This revised approach integrates feature fusion and inverse residual attention mechanisms. The aim of YOLO-IE is to improve the detection and classification of neuroblastoma tumors. In light of the image features, we have implemented the inverse residual-based attention structure (iRMB) within the detection network of YOLOv8, thereby enhancing the model's ability to focus on significant features present in the images. Additionally, we have incorporated the centered feature pyramid EVC module. Experimental results show that the proposed detection network, named YOLO-IE, attains a mean Average Precision (mAP) 7.9% higher than the baseline model, YOLO. The individual contributions of iRMB and EVC to the performance improvement are 0.8% and 3.6% above the baseline model, respectively. This study represents a significant advancement in the field, as it not only facilitates the detection and classification of neuroblastoma but also demonstrates the considerable potential of machine learning and artificial intelligence in the realm of medical diagnosis.

PMID:40287486 | DOI:10.1038/s41598-025-99451-5

Sci Data. 2025 Apr 26;12(1):703. doi: 10.1038/s41597-025-04965-2.

ABSTRACT

The performance and scope of computer vision methods applied to ophthalmic images is highly dependent on the availability of labelled training data. While there are a number of colour fundus photography datasets, FOVEA is to the best of our knowledge the first dataset that matches high-quality annotations in the intraoperative domain with those in the preoperative one. It comprises data from 40 patients collected at Moorfields Eye Hospital (London, UK) and includes preoperative and intraoperative retinal vessel and optic disc annotations performed by two independent clinical research fellows, as well as short video clips showing the retinal fundus though biomicroscopy imaging in the intraoperative setting. The annotations were validated and converted into binary segmentation masks, with the code used available on GitHub. We expect this data to be useful for deep learning applications aimed at supporting surgeons during vitreoretinal surgery procedures e.g. by localising points of interest or registering additional imaging modalities.

PMID:40287417 | DOI:10.1038/s41597-025-04965-2

Curr Probl Diagn Radiol. 2025 Apr 18:S0363-0188(25)00082-9. doi: 10.1067/j.cpradiol.2025.04.011. Online ahead of print.

ABSTRACT

PURPOSE: This review aims to provide a comprehensive overview of the transformative impact of FDA-approved artificial intelligence (AI) products in abdominal imaging. It explores the evolution of AI in radiology, its rigorous FDA clearance process, and its role in revolutionizing diagnostic and non-diagnostic tasks across various abdominal organs.

METHODS: Through a review of literature, this study categorizes AI products based on their applications in liver, prostate, bladder, kidney, and overall abdominal imaging. It analyzes the diagnostic and non-diagnostic functionalities of these AI solutions, elucidating their capabilities in enhancing disease detection, image quality, workflow efficiency, and longitudinal comparison standardization.

RESULTS: The review identifies numerous FDA-approved AI products tailored for abdominal imaging, showcasing their diverse applications, from lesion detection and characterization to volume estimation and quantification of organ health parameters. These AI solutions have demonstrated their efficacy in improving diagnostic accuracy, streamlining radiological workflows, and ultimately optimizing patient care across various abdominal pathologies.

CONCLUSION: In conclusion, the integration of AI into abdominal imaging represents a paradigm shift in modern radiology. By empowering radiologists with advanced tools for timely diagnosis, precise treatment planning, and improved patient outcomes, FDA-approved AI products herald a new era of innovation in abdominal imaging. Collaboration between developers, regulatory bodies, and the medical community will be paramount in harnessing the full potential of AI to reshape the future of abdominal radiology.

PMID:40287285 | DOI:10.1067/j.cpradiol.2025.04.011

Neuroscience. 2025 Apr 24:S0306-4522(25)00328-8. doi: 10.1016/j.neuroscience.2025.04.035. Online ahead of print.

ABSTRACT

Alzheimer's Disease (AD) as one of the most prevalent neurodegenerative disorders worldwide, characterized by significant memory and cognitive decline in its later stages, severely impacting daily lives. Consequently, early diagnosis and accurate assessment are crucial for delaying disease progression. In recent years, multimodal imaging has gained widespread adoption in AD diagnosis and research, particularly the combined use of Magnetic Resonance Imaging (MRI) and Positron Emission Tomography (PET). The complementarity of these modalities in structural and metabolic information offers a unique advantage for comprehensive disease understanding and precise diagnosis. With the rapid advancement of deep learning techniques, efficient fusion of MRI and PET multimodal data has emerged as a prominent research focus. This review systematically surveys the latest advancements in deep learning-based multimodal fusion of MRI and PET images for AD research, with a particular focus on studies published in the past five years (2021-2025). It first introduces the main sources of AD-related data, along with data preprocessing and feature extraction methods. Then it summarizes performance metrics and multimodal fusion techniques. Next, it explores the application of various deep learning models and their variants in multimodal fusion tasks. Finally, it analyzes the key challenges currently faced in the field, including data scarcity and imbalance, inter-institutional data heterogeneity, etc., and discusses potential solutions and future research directions. This review aims to provide systematic guidance for researchers in the field of MRI and PET multimodal fusion, with the ultimate goal of advancing the development of early AD diagnosis and intervention strategies.

PMID:40286904 | DOI:10.1016/j.neuroscience.2025.04.035

Jpn J Radiol. 2025 Apr 26. doi: 10.1007/s11604-025-01790-w. Online ahead of print.

ABSTRACT

PURPOSE: This study aimed to evaluate the efficacy of deep learning-based reconstruction (DLR) in improving pancreatic diffusion-weighted imaging (DWI) quality.

MATERIALS AND METHODS: In total, 117 patients (mean age of 68.0 ± 12.9 years) suspected of pancreatic diseases underwent magnetic resonance imaging (MRI) between July and December 2023. MRI sequences included respiratory-gated conventional diffusion-weighted images (RGC-DWIs), respiratory-gated diffusion-weighted images with deep learning-based reconstruction (DLR) (RGDLR-DWIs), and breath-hold diffusion-weighted images with DLR (BHDLR-DWIs) (short TE and long TE equal to other DWIs) at a 3 T MR system. Among these patients, 27 had solid lesions. Two radiologists qualitatively assessed pancreatic shape, main pancreatic duct (MPD) visualization, and solid lesion conspicuity using a 5-point scale. Quantitative analysis included apparent diffusion coefficient (ADC) values for pancreatic parenchyma and solid lesions, signal-to-noise ratio (SNR), pancreas-to-muscle signal-intensity ratio (PM-SIR) and lesion-to-pancreas signal-intensity ratio (LP-SIR). Differences among DWI sequences were analyzed using Friedman's and Bonferroni's tests.

RESULTS: Qualitatively, BHDLR-DWIs (short TE) had the highest scores for pancreatic shape and MPD but lowest for solid lesions visibility, whereas RGDLR-DWIs had the highest score for solid lesions. Quantitatively, BHDLR-DWIs (short TE) had the lowest ADC values for pancreatic parenchyma and solid lesions, with the highest PM-SIR. There was no significant difference between BHDLR-DWIs (short TE) and RGDLR-DWIs for solid lesion ADC values. RGC-DWIs had the highest SNR, though differences from RGDLR-DWIs and BHDLR-DWIs (short TE) were not significant. Although LP-SIR in RGDLR-DWIs were the lowest, the difference was not significant.

CONCLUSION: BHDLR-DWIs (short TE) provided the best pancreatic morphology image quality, whereas RGDLR-DWIs were superior for solid lesion detection.

PMID:40285832 | DOI:10.1007/s11604-025-01790-w

Abdom Radiol (NY). 2025 Apr 26. doi: 10.1007/s00261-025-04949-1. Online ahead of print.

ABSTRACT

PURPOSE: To establish and validate a model based on deep learning (DL), integrating radiomic features with relevant clinical features to generate nomogram, for predicting preoperative serosal invasion in gastric cancer (GC).

METHODS: This retrospective study included 335 patients from dual centers. T staging (T1-3 or T4) was used to assess serosal invasion. Radiomic features were extracted from primary GC lesions in the venous phase CT, and DL features from 8 transfer learning models were combined to create the Hand-crafted Radiomics and Deep Learning Radiomics (HCR-DLR) model. The Clinical (CL) model was built using clinical features, and both were combined into the Clinical and Radiomics Combined (CRC) model. In total, 15 predictive models were developed using 5 machine learning algorithms. The best-performing models were visualized as nomograms.

RESULTS: The total of 14 radiomic features, 13 DL features, and 2 clinical features were considered valuable through dimensionality reduction and selection. Among the constructed models: CRC model (AUC, training cohort: 0.9212; internal test cohort: 0.8743; external test cohort: 0.8853) than HCR-DLR model (AUC, training cohort: 0.8607; internal test cohort: 0.8543; external test cohort: 0.8824) and CL model (AUC, training cohort: 0.7632; internal test cohort: 0.7219; external test cohort: 0.7294) showed better performance. A nomogram based on the logistic CL model was drawn to facilitate the usage and showed its excellent predictive performance.

CONCLUSION: The predictive performance of the CRC Model, which integrates clinical features, radiomic features, and DL features, exhibits robust predictive capability and can serve as a simple, non-invasive, and practical tool for predicting the serosal invasion status of GC.

PMID:40285792 | DOI:10.1007/s00261-025-04949-1

Chem Res Toxicol. 2025 Apr 26. doi: 10.1021/acs.chemrestox.4c00560. Online ahead of print.

ABSTRACT

Transthyretin (TTR) plays a vital role in thyroid hormone transport and homeostasis in both the blood and target tissues. Interactions between exogenous compounds and TTR can disrupt the function of the endocrine system, potentially causing toxicity. In the Tox24 challenge, we leveraged the data set provided by the organizers to develop a deep learning-based consensus model, integrating sPhysNet, KANO, and GGAP-CPI for predicting TTR binding affinity. Each model utilized distinct levels of molecular information, including 2D topology, 3D geometry, and protein-ligand interactions. Our consensus model achieved favorable performance on the blind test set, yielding an RMSE of 20.8 and ranking fifth among all submissions. Following the release of the blind test set, we incorporated the leaderboard test set into our training data, further reducing the RMSE to 20.6 in an offlineretrospective study. These results demonstrate that combining three regression models across different modalities significantly enhances the predictive accuracy. Furthermore, we employ the standard deviation of the consensus model's ensemble outputs as an uncertainty estimate. Our analysis reveals that both the RMSE and interval error of predictions increase with rising uncertainty, indicating that the uncertainty can serve as a useful measure of prediction confidence. We believe that this consensus model can be a valuable resource for identifying potential TTR binders and predicting their binding affinity in silico. The source code for data preparation, model training, and prediction can be accessed at https://github.com/xiaolinpan/tox24_challenge_submission_yingkai_lab.

PMID:40285676 | DOI:10.1021/acs.chemrestox.4c00560

Adv Mater. 2025 Apr 26:e2419096. doi: 10.1002/adma.202419096. Online ahead of print.

ABSTRACT

Shear-stiffening materials, renowned for their rate-dependent behavior, hold immense potential for impact-resistant applications but are often constrained by limited load-bearing capacity under extreme conditions. In this study, a novel hybrid additive manufacturing strategy that successfully achieves anisotropic structural design of shear-stiffening materials is proposed. In this strategy, fused deposition modeling (FDM) is synergistically combined with direct ink writing (DIW) to fabricate lattice-structured soft-hard phase elastomer composites (TPR-SSE composites) with enhanced mechanical properties. Through quasistatic characterization and dynamic impact experiments, complemented by noncontact optical measurement and finite element simulation, the mechanical enhancement mechanisms imparted by the lattice architecture are systematically uncovered. The resulting composites exhibit exceptional load-bearing capacity under quasistatic conditions and superior energy dissipation under dynamic impacts, making them ideal for advanced protective systems. Building on this, smart sports shoes featuring a deep-learning-based smart sensing module that integrates structural customizability, buffering capacity, and gait recognition, are developed. This work provides a transformative structure design approach to shear-stiffening materials systems, paving the way for next-generation intelligent wearable protection applications.

PMID:40285578 | DOI:10.1002/adma.202419096

Adv Sci (Weinh). 2025 Apr 26:e2417682. doi: 10.1002/advs.202417682. Online ahead of print.

ABSTRACT

The increasing demand for automation in livestock farming scenarios highlights the need for effective noncontact measurement methods. The current methods typically require either fixed postures and specific positions of the target animals or high computational demands, making them difficult to implement in practical situations. In this study, a novel dual-network framework is presented that extracts accurate contour information instead of segmented images from unconstrained pigs and then directly employs this information to obtain precise liveweight estimates. The experimental results demonstrate that the developed framework achieves high accuracy, providing liveweight estimates with an R2 value of 0.993. When contour information is used directly to estimate the liveweight, the real-time performance of the framework can reach 1131.6 FPS. This achievement sets a new benchmark for accuracy and efficiency in non-contact liveweight estimation. Moreover, the framework holds significant practical value, equipping farmers with a robust and scalable tool for precision livestock management in dynamic, real-world farming environments. Additionally, the Liveweight and Instance Segmentation Annotation of Pigs dataset is introduced as a comprehensive resource designed to support further advancements and validation in this field.

PMID:40285549 | DOI:10.1002/advs.202417682

Brief Bioinform. 2025 Mar 4;26(2):bbaf186. doi: 10.1093/bib/bbaf186.

ABSTRACT

Deep learning (DL) methods have drastically advanced structure-based drug discovery by directly predicting protein structures from sequences. Recently, these methods have become increasingly accurate in predicting complexes formed by multiple protein chains. We evaluated these advancements to predict and accurately model the largest receptor family and its cognate peptide hormones. We benchmarked DL tools, including AlphaFold 2.3 (AF2), AlphaFold 3 (AF3), Chai-1, NeuralPLexer, RoseTTAFold-AllAtom, Peptriever, ESMFold, and D-SCRIPT, to predict interactions between G protein-coupled receptors (GPCRs) and their endogenous peptide ligands. Our results showed that structure-aware models outperformed language models in peptide binding classification, with the top-performing model achieving an area under the curve of 0.86 on a benchmark set of 124 ligands and 1240 decoys. Rescoring predicted structures on local interactions further improved the principal ligand discovery among decoy peptides, whereas DL-based approaches did not. We explored a competitive tournament approach for modeling multiple peptides simultaneously on a single GPCR, which accelerates the performance but reduces true-positive recovery. When evaluating the binding poses of 67 recent complexes, AF2 reproduced the correct binding modes in nearly all cases (94%), surpassing those of both AF3 and Chai-1. Confidence scores correlate with structural binding mode accuracy, which provides a guide for interpreting interface predictions. These results demonstrated that DL models can reliably rediscover peptide binders, aid peptide drug discovery, and guide the selection of optimal tools for GPCR-targeted therapies. To this end, we provided a practical guide for selecting the best models for specific applications and an independent benchmarking set for future model evaluation.

PMID:40285358 | DOI:10.1093/bib/bbaf186

Sensors (Basel). 2025 Apr 21;25(8):2623. doi: 10.3390/s25082623.

ABSTRACT

Automated storage and retrieval systems (AS/RS) play a crucial role in modern logistics, yet effectively monitoring cargo stacking patterns remains challenging. While computer vision and deep learning offer promising solutions, existing methods struggle to balance detection accuracy, computational efficiency, and environmental adaptability. This paper proposes a novel machine vision-based detection algorithm that integrates a pallet surface object detection network (STEGNet) with a box edge detection algorithm. STEGNet's core innovation is the Efficient Gated Pyramid Feature Network (EG-FPN), which integrates a Gated Feature Fusion module and a Lightweight Attention Mechanism to optimize feature extraction and fusion. In addition, we introduce a geometric constraint method for box edge detection and employ a Perspective-n-Point (PnP)-based 2D-to-3D transformation approach for precise pose estimation. Experimental results show that STEGNet achieves 93.49% mAP on our proposed GY Warehouse Box View 4-Dimension (GY-WSBW-4D) dataset and 83.2% mAP on the WSGID-B dataset, surpassing existing benchmarks. The lightweight variant maintains competitive accuracy while reducing the model size by 34% and increasing the inference speed by 68%. In practical applications, the system achieves pose estimation with a Mean Absolute Error within 4 cm and a Rotation Angle Error below 2°, demonstrating robust performance in complex warehouse environments. This research provides a reliable solution for automated cargo stack monitoring in modern logistics systems.

PMID:40285312 | DOI:10.3390/s25082623

Sensors (Basel). 2025 Apr 20;25(8):2615. doi: 10.3390/s25082615.

ABSTRACT

During image collection, images are often polluted by noise because of imaging conditions and equipment limitations. Images are also disturbed by external noise during compression and transmission, which adversely affects consequent processing, like image segmentation, target recognition, and text detection. A two-dimensional amplitude image is one of the most common image categories, which is widely used in people's daily life and work. Research on this kind of image-denoising algorithm is a hotspot in the field of image denoising. Conventional denoising methods mainly use the nonlocal self-similarity of images and sparser representatives in the converted domain for image denoising. In particular, the three-dimensional block matching filtering (BM3D) algorithm not only effectively removes the image noise but also better retains the detailed information in the image. As artificial intelligence develops, the deep learning-based image-denoising method has become an important research direction. This review provides a general overview and comparison of traditional image-denoising methods and deep neural network-based image-denoising methods. First, the essential framework of classic traditional denoising and deep neural network denoising approaches is presented, and the denoising approaches are classified and summarized. Then, existing denoising methods are compared with quantitative and qualitative analyses on a public denoising dataset. Finally, we point out some potential challenges and directions for future research in the field of image denoising. This review can help researchers clearly understand the differences between various image-denoising algorithms, which not only helps them to choose suitable algorithms or improve and innovate on this basis but also provides research ideas and directions for subsequent research in this field.

PMID:40285303 | DOI:10.3390/s25082615

Sensors (Basel). 2025 Apr 20;25(8):2608. doi: 10.3390/s25082608.

ABSTRACT

Deep learning has been widely applied in building segmentation from high-resolution remote sensing (HRS) images. However, HRS images suffer from insufficient complementary representation of target points in terms of capturing details and global information. To this end, we propose a novel building segmentation model for HRS images, termed C_ASegformer. Specifically, we design a Deep Layered Enhanced Fusion (DLEF) module to integrate hierarchical information from different receptive fields, thereby enhancing the feature representation capability of HRS information from global to detailed levels. Additionally, we introduce a Triplet Attention (TA) Module, which establishes dependency relationships between buildings and the environment through multi-directional rotation operations and residual transformations. Furthermore, we propose a Multi-Level Dilated Connection (MDC) Module to efficiently capture contextual relationships across different scales at a low computational cost. We conduct comparative experiments with several state-of-the-art models on three datasets, including the Massachusetts dataset, the INRIA dataset, and the WHU dataset. On the Massachusetts dataset, C_ASegformer achieves 95.42%, 85.69%, and 75.46% for OA, F1score, and mIoU, respectively. C_ASegformer shows more accurate performance, demonstrating the validity and sophistication of the model.

PMID:40285301 | DOI:10.3390/s25082608

Sensors (Basel). 2025 Apr 20;25(8):2603. doi: 10.3390/s25082603.

ABSTRACT

Challenges such as perspective distortion, irregular reading regions, and complex backgrounds in natural scenes hinder the accuracy and efficiency of automatic meter reading systems. Current mainstream approaches predominantly utilize object-detection-based methods without optimizing for text characteristics, while enhancements in detection robustness under complex backgrounds typically focus on data preprocessing rather than model architecture. To address these limitations, a novel end-to-end framework, i.e., EDPNet (Efficient DB and PARSeq Network), is proposed to integrate efficient boundary detection and text recognition. EDPNet comprises two key components, EDNet for detection and EPNet for recognition, where EDNet employs EfficientNetV2-s as its backbone with the Multi-Scale KeyDrop Attention (MSKA) and Efficient Multi-scale Attention (EMA) mechanisms to address perspective distortion and complex background challenges, respectively. During the recognition stage, EPNet integrates a DropKey Attention module into the PARSeq encoder, enhancing the recognition of irregular readings while effectively mitigating overfitting. Experimental evaluations show that EDNet achieves an F1-score of 0.997988, outperforming DBNet++ (ResNet50) by 0.61%. In challenging scenarios, EDPNet surpasses state-of-the-art methods by 0.7~1.9% while reducing parameters by 20.03%. EPNet achieves 90.0% recognition accuracy, exceeding the current best performance by 0.2%. The proposed framework delivers superior accuracy and robustness in challenging conditions while remaining lightweight.

PMID:40285298 | DOI:10.3390/s25082603

Clin Nucl Med. 2025 Apr 24. doi: 10.1097/RLU.0000000000005928. Online ahead of print.

ABSTRACT

PURPOSE: The aim of this study was to generate and validate artificial delayed-phase technetium-99m methoxyisobutylisonitrile scintigraphy (aMIBI) images from early-phase technetium-99m methoxyisobutylisonitrile scintigraphy (eMIBI) images.

PATIENTS AND METHODS: This retrospective study included patients with hyperparathyroidism who underwent dual-phase technetium-99m methoxyisobutylisonitrile (MIBI) scintigraphy at 2 centers. The patients were divided into a training set (n = 980), an internal test set (n = 100), and an external test set (n = 253). The generation of aMIBI images from eMIBI images was performed using an unpaired image-to-image translation method. Receiver operating characteristic curves and the area under the curves (AUCs) were used to evaluate the diagnostic performance of aMIBI and eMIBI images in identifying hyperfunctioning parathyroid lesions in both the internal and external test sets. In addition, an artificial intelligence (AI)-assisted diagnostic model combining aMIBI and clinical data was evaluated.

RESULTS: The AUCs of aMIBI images were significantly higher than those of eMIBI images (internal test set: 0.944 vs 0.658, P < 0.001; external test set: 0.900 vs 0.761, P < 0.001). The performance of the AI-assisted diagnostic models combining aMIBI images and clinical data was significantly better than those of the aMIBI-only models in both the internal (AUC: 0.974 vs 0.944, P = 0.020) and external (AUC: 0.953 vs 0.900, P < 0.001) test sets.

CONCLUSIONS: The diagnostic performance of aMIBI images in identifying hyperfunctioning parathyroid lesions was significantly superior to that of eMIBI images in patients with hyperparathyroidism. Models combining aMIBI images with clinical information enhanced the diagnostic performance even further.

PMID:40279678 | DOI:10.1097/RLU.0000000000005928

ACS Sens. 2025 Apr 25. doi: 10.1021/acssensors.5c00490. Online ahead of print.

ABSTRACT

As metabolic byproducts and representative indicators of food spoilage, the monitoring and detection for biogenic amines (BAs) are crucial but challenging for food quality assessment. Here, a strategy is proposed by combining fluorescent probe molecular engineering with a portable detection platform integrating a smartphone and a deep convolutional neural network (DCNN). Four ratiometric fluorescent probes with tunable intramolecular charge transfer (ICT) properties are designed by introducing different electron-withdrawing substituents (-F, -OCH3, -Py, and -CN) to the carbazole. Notably, CNCz exhibits the strongest ICT property and superior sensing performance, with a satisfying detection limit (11 ppb), rapid response (<5 s), and discriminative bathochromic shift (110 nm). Then, a smartphone-based detection platform is fabricated, which enables rapid, visual, and on-site quantitative evaluation of BAs. Furthermore, by integrating DCNN, this platform achieves an impressive 98.5% accuracy in predicting meat freshness. Hereby, this study not only provides a molecular engineering strategy to fine-tune the intrinsic ICT properties to gain high-performance ratiometric fluorescent probes but also presents an intelligent detection platform for BAs and meat freshness with high practical applicability.

PMID:40279659 | DOI:10.1021/acssensors.5c00490

JMIR Med Inform. 2025 Apr 25;13:e64963. doi: 10.2196/64963.

ABSTRACT

BACKGROUND: With the rapid development of artificial intelligence (AI) technology, especially generative AI, large language models (LLMs) have shown great potential in the medical field. Through massive medical data training, it can understand complex medical texts and can quickly analyze medical records and provide health counseling and diagnostic advice directly, especially in rare diseases. However, no study has yet compared and extensively discussed the diagnostic performance of LLMs with that of physicians.

OBJECTIVE: This study systematically reviewed the accuracy of LLMs in clinical diagnosis and provided reference for further clinical application.

METHODS: We conducted searches in CNKI (China National Knowledge Infrastructure), VIP Database, SinoMed, PubMed, Web of Science, Embase, and CINAHL (Cumulative Index to Nursing and Allied Health Literature) from January 1, 2017, to the present. A total of 2 reviewers independently screened the literature and extracted relevant information. The risk of bias was assessed using the Prediction Model Risk of Bias Assessment Tool (PROBAST), which evaluates both the risk of bias and the applicability of included studies.

RESULTS: A total of 30 studies involving 19 LLMs and a total of 4762 cases were included. The quality assessment indicated a high risk of bias in the majority of studies, primary cause is known case diagnosis. For the optimal model, the accuracy of the primary diagnosis ranged from 25% to 97.8%, while the triage accuracy ranged from 66.5% to 98%.

CONCLUSIONS: LLMs have demonstrated considerable diagnostic capabilities and significant potential for application across various clinical cases. Although their accuracy still falls short of that of clinical professionals, if used cautiously, they have the potential to become one of the best intelligent assistants in the field of human health care.

PMID:40279517 | DOI:10.2196/64963

J Phys Chem Lett. 2025 Apr 25:4463-4473. doi: 10.1021/acs.jpclett.5c00714. Online ahead of print.

ABSTRACT

We investigate the linear absorption spectra of the organic photovoltaic nonfullerene acceptor Y6 in chloroform using perturbative and nonperturbative approaches with atomistic details. Direct nonadiabatic semiclassical mapping dynamics reveal population and coherence evolution during and after ultrafast light pulse, revealing dominant absorption to the S1 state and subsequent oscillatory polarization. The simulated spectra accurately reproduce experimental peak positions and broadening, corresponding to transitions from the ground state to the S1, S2, and S6 excited states. Time-dependent radial distribution functions offer atomistic insights into solvent reorganization in response to charge redistribution. These findings enhance the understanding of nonadiabatic dynamics in Y6 and provide a consistent protocol for simulating electronic spectroscopy in condensed-phase systems.

PMID:40279488 | DOI:10.1021/acs.jpclett.5c00714

Bioinformatics. 2025 Apr 25:btaf203. doi: 10.1093/bioinformatics/btaf203. Online ahead of print.

ABSTRACT

SUMMARY: Differentiable folding is an emerging paradigm for RNA design in which a probabilistic sequence representation is optimized via gradient descent. However, given the significant memory overhead of differentiating the expected partition function over all RNA sequences, the existing proof-of-concept algorithm only scales to ≤50 nucleotides. We present JAX-RNAfold, an open-source software package for our drastically improved differentiable folding algorithm that scales to 1,250 nucleotides on a single GPU. Our software permits the natural inclusion of differentiable folding as a module in larger deep learning pipelines, as well as complex RNA design procedures such as mRNA design with flexible objective functions.

AVAILABILITY AND IMPLEMENTATION: JAX-RNAfold is hosted on GitHub (https://github.com/rkruegs123/jax-rnafold) and can be installed locally as a Python package. All source code is also archived on Zenodo (https://doi.org/10.5281/zenodo.15003072).

CONTACT: Please email max.ward@uwa.edu.au with any questions.

SUPPLEMENTARY INFORMATION: Please refer to the online-only Supplementary Material.

PMID:40279486 | DOI:10.1093/bioinformatics/btaf203

PLoS One. 2025 Apr 25;20(4):e0317514. doi: 10.1371/journal.pone.0317514. eCollection 2025.

ABSTRACT

Global awareness of sustainable development has heightened interest in green buildings as a key strategy for reducing energy consumption and carbon emissions. Accurate prediction of energy consumption plays a vital role in developing effective energy management and conservation strategies. This study addresses these challenges by proposing an advanced deep learning framework that integrates Time-Dependent Variational Autoencoder (TD-VAE) with Adaptive Gated Self-Attention GRU (AGSA-GRU). The framework incorporates self-attention mechanisms and Multi-Task Learning (MTL) strategies to capture long-term dependencies and complex patterns in energy consumption time series data, while simultaneously optimizing prediction accuracy and anomaly detection. Experiments on two public green building energy consumption datasets validate the effectiveness of our proposed approach. Our method achieves a prediction accuracy of 93.2%, significantly outperforming traditional deep learning methods and existing techniques. ROC curve analysis demonstrates our model's robustness, achieving an Area Under the Curve (AUC) of 0.91 while maintaining a low false positive rate (FPR) and high true positive rate (TPR). This study presents an efficient solution for green building energy consumption prediction, contributing significantly to energy conservation, emission reduction, and sustainable development in the construction industry.

PMID:40279377 | DOI:10.1371/journal.pone.0317514