Discov Oncol. 2025 Feb 12;16(1):165. doi: 10.1007/s12672-025-01942-4.

ABSTRACT

PURPOSE: In this work, we propose to develop a 3D U-Net-based deep learning model that accurately predicts the dose distribution for breast cancer radiotherapy.

METHODS: This study included 176 breast cancer patients, divided into training, validating and testing sets. A deep learning model based on the 3D U-Net architecture was developed to predict dose distribution, which employed a double encoder combination attention (DECA) module, a cross stage partial + Resnet + Attention (CRA) module, a difficulty perception and a critical regions loss. The performance and generalization ability of this model were evaluated by the voxel mean absolute error (MAE), several clinically relevant dosimetric indexes and 3D gamma passing rates.

RESULTS: Our model accurately predicted the 3D dose distributions with each dosage level mirroring the clinical reality in shape. The generated dose-volume histogram (DVH) matched with the ground truth curve. The total dose error of our model was below 1.16 Gy, complying with clinical usage standards. When compared to other exceptional models, our model optimally predicted eight out of nine regions, and the prediction errors for the first planning target volume (PTV1) and PTV2 were merely 1.03 Gy and 0.74 Gy. Moreover, the mean 3%/3 mm 3D gamma passing rates for PTV1, PTV2, Heart and Lung L achieved 91.8%, 96.4%, 91.5%, and 93.2%, respectively, surpassing the other models and meeting clinical standards.

CONCLUSIONS: This study developed a new deep learning model based on 3D U-Net that can accurately predict dose distributions for breast cancer radiotherapy, which can improve the quality and planning efficiency.

PMID:39937302 | DOI:10.1007/s12672-025-01942-4

Eur Radiol. 2025 Feb 12. doi: 10.1007/s00330-025-11401-x. Online ahead of print.

ABSTRACT

OBJECTIVE: To systematically evaluate glioma radiomics literature on differentiating between radiation-induced brain injury and tumor recurrence.

METHODS: Literature was searched on PubMed and Web of Science (end date: May 7, 2024). Quality of eligible papers was assessed using METhodological RadiomICs Score (METRICS) and Radiomics Quality Score (RQS). Reliability of quality scoring tools were analyzed. Meta-analysis, meta-regression, and subgroup analysis were performed.

RESULTS: Twenty-seven papers were included in the qualitative assessment. Mean average METRICS score and RQS percentage score across three readers was 57% (SD, 14%) and 16% (SD, 12%), respectively. Score-wise inter-rater agreement for METRICS ranged from poor to excellent, while RQS demonstrated moderate to excellent agreement. Item-wise agreement was moderate for both tools. Meta-analysis of 11 eligible studies yielded an estimated area under the receiver operating characteristic curve of 0.832 (95% CI, 0.757-0.908), with significant heterogeneity (I2 = 91%) and no statistical publication bias (p = 0.051). Meta-regression did not identify potential sources of heterogeneity. Subgroup analysis revealed high heterogeneity across all subgroups, with the lowest I2 at 68% in studies with proper validation and higher quality scores. Statistical publication bias was generally not significant, except in the subgroup with the lowest heterogeneity (p = 0.044). However, most studies in both qualitative analysis (26/27; 96%) and primary meta-analysis (10/11; 91%) reported positive effects of radiomics, indicating high non-statistical publication bias.

CONCLUSION: While a good performance was noted for radiomics, results should be interpreted cautiously due to heterogeneity, publication bias, and quality issues thoroughly examined in this study.

KEY POINTS: Question Radiomic literature on distinguishing radiation-induced brain injury from glioma recurrence lacks systematic reviews and meta-analyses that assess methodological quality using radiomics-specific tools. Findings While the results are encouraging, there was substantial heterogeneity, publication bias toward positive findings, and notable concerns regarding methodological quality. Clinical relevance Meta-analysis results need cautious interpretation due to significant problems detected during the analysis (e.g., suboptimal quality, heterogeneity, bias), which may help explain why radiomics has not yet been translated into clinical practice.

PMID:39937273 | DOI:10.1007/s00330-025-11401-x

Abdom Radiol (NY). 2025 Feb 12. doi: 10.1007/s00261-025-04834-x. Online ahead of print.

ABSTRACT

The purpose of this study was to explore the association between structural changes in abdominal organs and visceral fat obesity (VFO) using a fully automated three-dimensional (3D) volumetric computed tomography (CT) measurement method based on deep learning algorithm. A total of 610 patients (295 men and 315 women; mean age, 68.4 years old) were included. Fully automated 3D volumetric CT measurements of the abdominal organs were performed to determine the volume and average CT attenuation values of each organ. All patients were divided into 2 groups based on the measured visceral fat area: the VFO group (≥ 100 cm2) and non-VFO group (< 100 cm2), and the structural changes in abdominal organs were compared between these groups. The volumes of all organs were significantly higher in the VFO group than in the non-VFO group (all of p < 0.001). Conversely, the CT attenuation values of all organs in the VFO group were significantly lower than those in the non-VFO group (all of p < 0.001). Pancreatic CT values (r = - 0.701, p < 0.001) were most strongly associated with the visceral fat, followed by renal CT values (r = - 0.525, p < 0.001) and hepatic CT values (r = - 0.510, p < 0.001). Fully automated 3D volumetric CT measurement using a deep learning algorithm has the potential to detect the structural changes in the abdominal organs, especially the pancreas, such as an increase in the volumes and a decrease in CT attenuation values, probably due to increased ectopic fat accumulation in patients with VFO. This technique may provide valuable imaging support for the early detection and intervention of metabolic-related diseases.

PMID:39937214 | DOI:10.1007/s00261-025-04834-x

Biomacromolecules. 2025 Feb 12. doi: 10.1021/acs.biomac.4c01243. Online ahead of print.

ABSTRACT

Natural topological proteins exhibit unique properties including enhanced stability, controlled quaternary structures, and dynamic switching properties, highlighting topology as a unique dimension in protein engineering. Although artificial design and synthesis of topological proteins have achieved certain success, their diversity and complexity remain rather limited due to the scarcity of available entangling motifs essential for the construction of nontrivial protein topologies. In this work, we developed a deep-learning model to predict the entanglement features of a homodimer based solely on its amino acid sequence via the Gauss linking number matrices. The model achieved a search speed that was dozens of times faster than AlphaFold-Multimer, while maintaining comparable mean squared error. It was used to screen for entangling motifs from the genome of a hyperthermophilic archaeon. We demonstrated the effectiveness of our model by successful wet-lab synthesis of protein catenanes using two candidate entangling motifs. These findings show the great potential of our model for advancing the design and synthesis of novel topological proteins.

PMID:39937127 | DOI:10.1021/acs.biomac.4c01243

J Med Eng Technol. 2025 Feb 12:1-14. doi: 10.1080/03091902.2025.2462310. Online ahead of print.

ABSTRACT

This research aimed to develop an algorithm for classifying scalogram images generated from electromyography data of patients with Rheumatoid Arthritis and Prolapsed Intervertebral Disc. Electromyography is valuable for assessing muscle function and diagnosing neurological disorders, but limitations, such as background noise, cross-talk, and inter-subject variability complicate the interpretation and assessment. To mitigate this, the present study uses scalogram images and attention-network architecture. The algorithm utilises a combination of features extracted from an attention module and a convolution feature module, followed by classification using a Convolutional Neural Network classifier. A comparison of eight alternative architectures, including individual implementations of attention and convolution filters and a Convolutional Neural Network-only model, shows that the hybrid Convolutional Neural Network model proposed in this study outperforms the others. The model exhibits excellent discriminatory ability between gait abnormalities with an accuracy of 96.7%, a precision of 95.2%, a recall of 94.8%, and an Area Under Curve of 0.99. These findings suggest that the proposed model is highly accurate in classifying scalogram images of electromyography signals and may have significant clinical implications for early diagnosis and treatment planning.

PMID:39936825 | DOI:10.1080/03091902.2025.2462310

Rev Soc Bras Med Trop. 2025 Feb 7;58:e004012024. doi: 10.1590/0037-8682-0178-2024. eCollection 2025.

ABSTRACT

BACKGROUND: Emerging infections have posed persistent threats to humanity throughout history. Rapid and unprecedented anthropogenic, behavioral, and social transformations witnessed in the past century have expedited the emergence of novel pathogens, intensifying their impact on the global human population.

METHODS: This study aimed to comprehensively analyze and compare the genomic sequences of four distinct viruses: SARS-CoV-2, SARS, MERS, and Ebola. Advanced genomic sequencing techniques and a Gated Recurrent Unit-based deep learning model were used to examine the intricate genetic makeup of these viruses. The proposed study sheds light on their evolutionary dynamics, transmission patterns, and pathogenicity and contributes to the development of effective diagnostic and therapeutic interventions.

RESULTS: This model exhibited exceptional performance as evidenced by accuracy values of 99.01%, 98.91%, 98.35%, and 98.04% for SARS-CoV-2, SARS, MERS, and Ebola respectively. Precision values ranged from 98.1% to 98.72%, recall values consistently surpassed 92%, and F1 scores ranged from 95.47% to 96.37%.

CONCLUSIONS: These results underscore the robustness of this model and its potential utility in genomic analysis, paving the way for enhanced understanding, preparedness, and response to emerging viral threats. In the future, this research will focus on creating better diagnostic instruments for the early identification of viral illnesses, developing vaccinations, and tailoring treatments based on the genetic composition and evolutionary patterns of different viruses. This model can be modified to examine a more extensive variety of diseases and recently discovered viruses to predict future outbreaks and their effects on global health.

PMID:39936709 | DOI:10.1590/0037-8682-0178-2024

Curr Neuropharmacol. 2025 Feb 6. doi: 10.2174/011570159X327908241121062335. Online ahead of print.

ABSTRACT

Neurodegenerative diseases represent a prevalent category of age-associated diseases. As human lifespans extend and societies become increasingly aged, neurodegenerative diseases pose a growing threat to public health. The lack of effective therapeutic drugs for both common and rare neurodegenerative diseases amplifies the medical challenges they present. Current treatments for these diseases primarily offer symptomatic relief rather than a cure, underscoring the pressing need to develop efficacious therapeutic interventions. Drug repositioning, an innovative and data-driven approach to research and development, proposes the re-evaluation of existing drugs for potential application in new therapeutic areas. Fueled by rapid advancements in artificial intelligence and the burgeoning accumulation of medical data, drug repositioning has emerged as a promising pathway for drug discovery. This review comprehensively examines drug repositioning for neurodegenerative diseases through the lens of translational informatics, encompassing data sources, computational models, and clinical applications. Initially, we systematized drug repositioning-related databases and online platforms, focusing on data resource management and standardization. Subsequently, we classify computational models for drug repositioning from the perspectives of drug-drug, drug-target, and drug-disease interactions into categories such as machine learning, deep learning, and networkbased approaches. Lastly, we highlight computational models presently utilized in neurodegenerative disease research and identify databases that hold potential for future drug repositioning efforts. In the artificial intelligence era, drug repositioning, as a data-driven strategy, offers a promising avenue for developing treatments suited to the complex and multifaceted nature of neurodegenerative diseases. These advancements could furnish patients with more rapid, cost-effective therapeutic options.

PMID:39936420 | DOI:10.2174/011570159X327908241121062335

Hum Brain Mapp. 2025 Feb 15;46(3):e70141. doi: 10.1002/hbm.70141.

ABSTRACT

Disorders of the central nervous system, including neurodegenerative diseases, frequently affect the brainstem and can present with focal atrophy. This study aimed to (1) optimize deep learning-based brainstem segmentation for a wide range of pathologies and T1-weighted image acquisition parameters, (2) conduct a systematic technical and clinical validation, (3) improve segmentation quality in the presence of brainstem lesions, and (4) make an optimized brainstem segmentation tool available for public use. An intentionally heterogeneous ground truth dataset (n = 257) was employed in the training of deep learning models based on multi-dimensional gated recurrent units (MD-GRU) or the nnU-Net method. Segmentation performance was evaluated against ground truth labels. FreeSurfer was used for benchmarking in subsequent validation. Technical validation, including scan-rescan repeatability (n = 46) and inter-scanner reproducibility (n = 20, 3 different scanners) in unseen data, was conducted in patients with cerebral small vessel disease. Clinical validation in unseen data was performed in 1-year follow-up data of 16 patients with multiple system atrophy, evaluating the annual percentage volume change. Two lesion filling algorithms were investigated to improve segmentation performance in 23 patients with multiple sclerosis. The MD-GRU and nnU-Net models demonstrated very good segmentation performance (median Dice coefficients ≥ 0.95 each) and outperformed a previously published model trained on a narrower dataset. Scan-rescan repeatability and inter-scanner reproducibility yielded similar Bland-Altman derived limits of agreement for longitudinal FreeSurfer (total brainstem volume repeatability/reproducibility 0.68/1.85), MD-GRU (0.72/1.46), and nnU-Net (0.48/1.52). All methods showed comparable performance in the detection of atrophy in the total brainstem (atrophy detected in 100% of patients) and its substructures. In patients with multiple sclerosis, lesion filling further improved the accuracy of brainstem segmentation. We enhanced and systematically validated two fully automated deep learning brainstem segmentation methods and released them publicly. This enables a broader evaluation of brainstem volume as a candidate biomarker for neurodegeneration.

PMID:39936343 | DOI:10.1002/hbm.70141

J Sci Food Agric. 2025 Feb 12. doi: 10.1002/jsfa.14177. Online ahead of print.

ABSTRACT

BACKGROUND: Silkworm pupae (SP), the pupal stage of an edible insect, have strong potential in the food, medicine, and cosmetic industries. Sex sorting is essential to enhance nutritional content and genetic traits in SP crossbreeding but it remains labor intensive and time consuming. An intelligent method is needed urgently to improve efficiency and productivity.

RESULTS: To address the problem, an automatic SP sex-separation system was developed based on computer vision and deep learning. Specifically, based on gonad features, a novel real-time SP sex identification model with cascaded spatial channel attention (CSCA) and G-GhostNet (GPU-Ghost Network) was developed, which can capture regions of interest and achieve feature diversity efficiently. A new loss function was proposed to reduce model complexity and avoid overfitting in the training. In comparison with benchmark methods on the test set, the new model achieved superior performance with an accuracy of 96.48%. The experimental sorting accuracy for SP reached 95.59%, validating the effectiveness of the novel gender-separation strategy.

CONCLUSION: This research presents a practical method for online SP gender separation, potentially aiding the production of high-quality SP. © 2025 Society of Chemical Industry.

PMID:39936219 | DOI:10.1002/jsfa.14177

Front Plant Sci. 2025 Jan 28;16:1523552. doi: 10.3389/fpls.2025.1523552. eCollection 2025.

ABSTRACT

With the development of deep learning technology and the widespread application of drones in the agricultural sector, the use of computer vision technology for target detection of pineapples has gradually been recognized as one of the key methods for estimating pineapple yield. When images of pineapple fields are captured by drones, the fruits are often obscured by the pineapple leaf crowns due to their appearance and planting characteristics. Additionally, the background in pineapple fields is relatively complex, and current mainstream target detection algorithms are known to perform poorly in detecting small targets under occlusion conditions in such complex backgrounds. To address these issues, an improved YOLOv8 target detection algorithm, named DPD-YOLO (Dense-Pineapple-Detection YOU Only Look Once), has been proposed for the detection of pineapples in complex environments. The DPD-YOLO model is based on YOLOv8 and introduces the attention mechanism (Coordinate Attention) to enhance the network's ability to extract features of pineapples in complex backgrounds. Furthermore, the small target detection layer has been fused with BiFPN (Bi-directional Feature Pyramid Network) to strengthen the integration of multi-scale features and enrich the extraction of semantic features. At the same time, the original YOLOv8 detection head has been replaced by the RT-DETR detection head, which incorporates Cross-Attention and Self-Attention mechanisms that improve the model's detection accuracy. Additionally, Focaler-IoU has been employed to improve CIoU, allowing the network to focus more on small targets. Finally, high-resolution images of the pineapple fields were captured using drones to create a dataset, and extensive experiments were conducted. The results indicate that, compared to existing mainstream target detection models, the proposed DPD-YOLO demonstrated superior detection performance for pineapples in situations where the background is complex and the targets are occluded. The mAP@0.5 reached 62.0%, representing an improvement of 6.6% over the original YOLOv8 algorithm, Precision increased by 2.7%, Recall improved by 13%, and F1-score rose by 10.3%.

PMID:39935949 | PMC:PMC11810954 | DOI:10.3389/fpls.2025.1523552

Front Oncol. 2025 Jan 28;15:1474590. doi: 10.3389/fonc.2025.1474590. eCollection 2025.

ABSTRACT

PURPOSE: This work investigates the use of a spherical projection-based U-Net (SPU-Net) segmentation model to improve meningioma segmentation performance and allow for uncertainty quantification.

METHODS: A total of 76 supratentorial meningioma patients treated with radiotherapy were studied. Gross tumor volumes (GTVs) were contoured by a single experienced radiation oncologist on high-resolution contrast-enhanced T1 MRI scans (T1ce), and both T1 and T1ce images were utilized for segmentation. SPU-Net, an adaptation of U-Net incorporating spherical image projection to map 2D images onto a spherical surface, was proposed. As an equivalence of a nonlinear image transform, projections enhance locoregional details while maintaining the global field of view. By employing multiple projection centers, SPU-Net generates various GTV segmentation predictions, the variance indicating the model's uncertainty. This uncertainty is quantified on a pixel-wise basis using entropy calculations and aggregated through Otsu's method for a final segmentation.

RESULTS/CONCLUSION: The SPU-Net model poses an advantage over traditional U-Net models by providing a quantitative method of displaying segmentation uncertainty. Regarding segmentation performance, SPU-Net demonstrated comparable results to a traditional U-Net in sensitivity (0.758 vs. 0.746), Dice similarity coefficient (0.760 vs. 0.742), reduced mean Hausdorff distance (mHD) (0.612 cm vs 0.744 cm), and reduced 95% Hausdorff distance (HD95) (2.682 cm vs 2.912 cm). SPU-Net not only is comparable to U-Net in segmentation performance but also offers a significant advantage by providing uncertainty quantification. The added SPU-Net uncertainty mapping revealed low uncertainty in accurate segments (e.g., within GTV or healthy tissue) and higher uncertainty in problematic areas (e.g., GTV boundaries, dural tail), providing valuable insights for potential manual corrections. This advancement is particularly valuable given the complex extra-axial nature of meningiomas and involvement with dural tissue. The capability to quantify uncertainty makes SPU-Net a more advanced and informative tool for segmentation, without sacrificing performance.

PMID:39935829 | PMC:PMC11810883 | DOI:10.3389/fonc.2025.1474590

Front Oncol. 2025 Jan 28;15:1528654. doi: 10.3389/fonc.2025.1528654. eCollection 2025.

ABSTRACT

OBJECTIVES: To develop a novel automatic delineation model, the Multi-Scale Channel Attention U-Net (MCAU-Net) model, for gallbladder segmentation on CT images of patients with liver cancer.

METHODS: We retrospectively collected the CT images from 120 patients with liver cancer, based on which ground truth was manually delineated by physicians. The images and ground truth constitute a dataset, which was proportionally divided into a training set (54%), a validation set (6%), and a test set (40%). Data augmentation was performed on the training set. Our proposed MCAU-Net model was employed for gallbladder segmentation and its performance was evaluated using Dice Similarity Coefficient (DSC), Jaccard Similarity Coefficient (JSC), Positive Predictive Value (PPV), Sensitivity (SE), Hausdorff Distance (HD), Relative Volume Difference (RVD), and Volumetric Overlap Error (VOE) metrics.

RESULTS: On the test set, MCAU-Net achieved DSC, JSC, PPV, SE, HD, RVD, and VOE values of 0.85 ± 0.22, 0.79 ± 0.23, 0.92 ± 0.14, 0.84 ± 0.23, 2.75 ± 0.98, 0.18 ± 0.48, and 0.22 ± 0.42, respectively. Compared to the control models, U-Net, SEU-Net and TransUNet, the MCAU-Net improved DSC 0.06, 0.04 and 0.06, JSC by 0.09, 0.06 and 0.09, PPV by 0.08, 0.08 and 0.05, SE by 0.05,0.05 and 0.07, and reduced HD by 0.45, 0.28 and 0.41, RVD by 0.07, 0.03 and 0.07, VOE by 0.04, 0.02 and 0.08 respectively. Qualitative results revealed that MCAU-Net produced smoother and more accurate boundaries, closer to the expert delineation, with less over-segmentation and under-segmentation and improved robustness.

CONCLUSIONS: The MCAU-Net model significantly improves gallbladder segmentation on CT images. It satisfies clinical requirements and enhances the efficiency of physicians, particularly in segmenting complex anatomical structures.

PMID:39935828 | PMC:PMC11810919 | DOI:10.3389/fonc.2025.1528654

Front Med (Lausanne). 2025 Jan 28;12:1492709. doi: 10.3389/fmed.2025.1492709. eCollection 2025.

ABSTRACT

OBJECTIVE: To systematically map the knowledge landscape and development trends in artificial intelligence (AI) applications for antimicrobial resistance (AMR) research through bibliometric analysis, providing evidence-based insights to guide future research directions and inform strategic decision-making in this dynamic field.

METHODS: A comprehensive bibliometric analysis was performed using the Web of Science Core Collection database for publications from 2014 to 2024. The analysis integrated multiple bibliometric approaches: VOSviewer for visualization of collaboration networks and research clusters, CiteSpace for temporal evolution analysis, and quantitative analysis of publication metrics. Key bibliometric indicators including co-authorship patterns, keyword co-occurrence, and citation impact were analyzed to delineate research evolution and collaboration patterns in this domain.

RESULTS: A collection of 2,408 publications was analyzed, demonstrating significant annual growth with publications increasing from 4 in 2014 to 549 in 2023 (22.7% of total output). The United States (707), China (581), and India (233) were the leading contributors in international collaborations. The Chinese Academy of Sciences (53), Harvard Medical School (43), and University of California San Diego (26) were identified as top contributing institutions. Citation analysis highlighted two major breakthroughs: AlphaFold's protein structure prediction (6,811 citations) and deep learning approaches to antibiotic discovery (4,784 citations). Keyword analysis identified six enduring research clusters from 2014 to 2024: sepsis, artificial neural networks, antimicrobial resistance, antimicrobial peptides, drug repurposing, and molecular docking, demonstrating the sustained integration of AI in antimicrobial therapy development. Recent trends show increasing application of AI technologies in traditional approaches, particularly in MALDI-TOF MS for pathogen identification and graph neural networks for large-scale molecular screening.

CONCLUSION: This bibliometric analysis shows the importance of artificial intelligence in enhancing the progress in the discovery of antimicrobial drugs especially toward the fight against AMR. From enhancing the fast, efficient and predictive performance of drug discovery methods, current AI capabilities have revealed observable potential to be proactive in combating the ever-growing challenge of AMR worldwide. This study serves not only an identification of current trends, but also, and especially, offers a strategic approach to further investigations.

PMID:39935800 | PMC:PMC11810743 | DOI:10.3389/fmed.2025.1492709

Front Cell Dev Biol. 2025 Jan 28;12:1517240. doi: 10.3389/fcell.2024.1517240. eCollection 2024.

ABSTRACT

PURPOSE: Using deep learning model to observe the blinking characteristics and evaluate the changes and their correlation with tear film characteristics in children with long-term use of orthokeratology (ortho-K).

METHODS: 31 children (58 eyes) who had used ortho-K for more than 1 year and 31 age and gender-matched controls were selected for follow-up in our ophthalmology clinic from 2021/09 to 2023/10 in this retrospective case-control study. Both groups underwent comprehensive ophthalmological examinations, including Ocular Surface Disease Index (OSDI) scoring, Keratograph 5M, and LipiView. A deep learning system based on U-Net and Swim-Transformer was proposed for the observation of blinking characteristics. The frequency of incomplete blinks (IB), complete blinks (CB) and incomplete blinking rate (IBR) within 20 s, as well as the duration of the closing, closed, and opening phases in the blink wave were calculated by our deep learning system. Relative IPH% was proposed and defined as the ratio of the mean of IPH% within 20 s to the maximum value of IPH% to indicate the extent of incomplete blinking. Furthermore, the accuracy, precision, sensitivity, specificity, F1 score of the overall U-Net-Swin-Transformer model, and its consistency with built-in algorithm were evaluated as well. Independent t-test and Mann-Whitney test was used to analyze the blinking patterns and tear film characteristics between the long-term ortho-K wearer group and the control group. Spearman's rank correlation was used to analyze the relationship between blinking patterns and tear film stability.

RESULTS: Our deep learning system demonstrated high performance (accuracy = 98.13%, precision = 96.46%, sensitivity = 98.10%, specificity = 98.10%, F1 score = 0.9727) in the observation of blinking patterns. The OSDI scores, conjunctival redness, lipid layer thickness (LLT), and tear meniscus height did not change significantly between two groups. Notably, the ortho-K group exhibited shorter first (11.75 ± 7.42 s vs. 14.87 ± 7.93 s, p = 0.030) and average non-invasive tear break-up times (NIBUT) (13.67 ± 7.0 s vs. 16.60 ± 7.24 s, p = 0.029) compared to the control group. They demonstrated a higher IB (4.26 ± 2.98 vs. 2.36 ± 2.55, p < 0.001), IBR (0.81 ± 0.28 vs. 0.46 ± 0.39, p < 0.001), relative IPH% (0.3229 ± 0.1539 vs. 0.2233 ± 0.1960, p = 0.004) and prolonged eye-closing phase (0.18 ± 0.08 s vs. 0.15 ± 0.07 s, p = 0.032) and opening phase (0.35 ± 0.12 s vs. 0.28 ± 0.14 s, p = 0.015) compared to controls. In addition, Spearman's correlation analysis revealed a negative correlation between incomplete blinks and NIBUT (for first-NIBUT, r = -0.292, p = 0.004; for avg-NIBUT, r = -0.3512, p < 0.001) in children with long-term use of ortho-K.

CONCLUSION: The deep learning system based on U-net and Swim-Transformer achieved optimal performance in the observation of blinking characteristics. Children with long-term use of ortho-K presented an increase in the frequency and rate of incomplete blinks and prolonged eye closing phase and opening phase. The increased frequency of incomplete blinks was associated with decreased tear film stability, indicating the importance of monitoring children's blinking patterns as well as tear film status in clinical follow-up.

PMID:39935789 | PMC:PMC11811098 | DOI:10.3389/fcell.2024.1517240

J R Stat Soc Series B Stat Methodol. 2024 Aug 5;87(1):98-131. doi: 10.1093/jrsssb/qkae082. eCollection 2025 Feb.

ABSTRACT

While fiducial inference was widely considered a big blunder by R.A. Fisher, the goal he initially set-'inferring the uncertainty of model parameters on the basis of observations'-has been continually pursued by many statisticians. To this end, we develop a new statistical inference method called extended Fiducial inference (EFI). The new method achieves the goal of fiducial inference by leveraging advanced statistical computing techniques while remaining scalable for big data. Extended Fiducial inference involves jointly imputing random errors realized in observations using stochastic gradient Markov chain Monte Carlo and estimating the inverse function using a sparse deep neural network (DNN). The consistency of the sparse DNN estimator ensures that the uncertainty embedded in observations is properly propagated to model parameters through the estimated inverse function, thereby validating downstream statistical inference. Compared to frequentist and Bayesian methods, EFI offers significant advantages in parameter estimation and hypothesis testing. Specifically, EFI provides higher fidelity in parameter estimation, especially when outliers are present in the observations; and eliminates the need for theoretical reference distributions in hypothesis testing, thereby automating the statistical inference process. Extended Fiducial inference also provides an innovative framework for semisupervised learning.

PMID:39935678 | PMC:PMC11809222 | DOI:10.1093/jrsssb/qkae082

Front Psychiatry. 2025 Jan 28;16:1508772. doi: 10.3389/fpsyt.2025.1508772. eCollection 2025.

ABSTRACT

INTRODUCTION: Depression is a serious mental health disease. Traditional scale-based depression diagnosis methods often have problems of strong subjectivity and high misdiagnosis rate, so it is particularly important to develop automatic diagnostic tools based on objective indicators.

METHODS: This study proposes a deep learning method that fuses multimodal data to automatically diagnose depression using facial video and audio data. We use spatiotemporal attention module to enhance the extraction of visual features and combine the Graph Convolutional Network (GCN) and the Long and Short Term Memory (LSTM) to analyze the audio features. Through the multi-modal feature fusion, the model can effectively capture different feature patterns related to depression.

RESULTS: We conduct extensive experiments on the publicly available clinical dataset, the Extended Distress Analysis Interview Corpus (E-DAIC). The experimental results show that we achieve robust accuracy on the E-DAIC dataset, with a Mean Absolute Error (MAE) of 3.51 in estimating PHQ-8 scores from recorded interviews.

DISCUSSION: Compared with existing methods, our model shows excellent performance in multi-modal information fusion, which is suitable for early evaluation of depression.

PMID:39935533 | PMC:PMC11811426 | DOI:10.3389/fpsyt.2025.1508772

J Multidiscip Healthc. 2025 Feb 7;18:675-695. doi: 10.2147/JMDH.S493873. eCollection 2025.

ABSTRACT

PURPOSE: Breast cancer is the most common major public health problems of women in the world. Until now, analyzing mammogram images is still the main method used by doctors to diagnose and detect breast cancers. However, this process usually depends on the experience of radiologists and is always very time consuming.

PATIENTS AND METHODS: We propose to introduce deep learning technology into the process for the facilitation of computer-aided diagnosis (CAD), and address the challenges of class imbalance, enhance the detection of small masses and multiple targets, and reduce false positives and negatives in mammogram analysis. Therefore, we adopted and enhanced RetinaNet to detect masses in mammogram images. Specifically, we introduced a novel modification to the network structure, where the feature map M5 is processed by the ReLU function prior to the original convolution kernel. This strategic adjustment was designed to prevent the loss of resolution for small mass features. Additionally, we introduced transfer learning techniques into training process through leveraging pre-trained weights from other RetinaNet applications, and fine-tuned our improved model using the INbreast dataset.

RESULTS: The aforementioned innovations facilitate superior performance of the enhanced RetiaNet model on the public dataset INbreast, as evidenced by a mAP (mean average precision) of 1.0000 and TPR (true positive rate) of 1.00 at 0.00 FPPI (false positive per image) on the INbreast dataset.

CONCLUSION: The experimental results demonstrate that our enhanced RetinaNet model defeats the existing models by having more generalization performance than other published studies, and it can also be applied to other types of patients to assist doctors in making a proper diagnosis.

PMID:39935433 | PMC:PMC11812562 | DOI:10.2147/JMDH.S493873

Sci Rep. 2025 Feb 11;15(1):5093. doi: 10.1038/s41598-025-89096-9.

ABSTRACT

Medical image segmentation plays a crucial role in addressing emerging healthcare challenges. Although several impressive deep learning architectures based on convolutional neural networks (CNNs) and Transformers have recently demonstrated remarkable performance, there is still potential for further performance improvement due to their inherent limitations in capturing feature correlations of input data. To address this issue, this paper proposes a novel encoder-decoder architecture called MedFuseNet that aims to fuse local and global deep feature representations with hybrid attention mechanisms for medical image segmentation. More specifically, the proposed approach contains two branches for feature learning in parallel: one leverages CNNs to learn local correlations of input data, and the other utilizes Swin-Transformer to capture global contextual correlations of input data. For feature fusion and enhancement, the designed hybrid attention mechanisms combine four different attention modules: (1) an atrous spatial pyramid pooling (ASPP) module for the CNN branch, (2) a cross attention module in the encoder for fusing local and global features, (3) an adaptive cross attention (ACA) module in skip connections for further performing fusion, and (4) a squeeze-and-excitation attention (SE-attention) module in the decoder for highlighting informative features. We evaluate our proposed approach on the public ACDC and Synapse datasets, and achieves the average DSC of 89.73% and 78.40%, respectively. Experimental results on these two datasets demonstrate the effectiveness of our proposed approach on medical image segmentation tasks, outperforming other used state-of-the-art approaches.

PMID:39934248 | DOI:10.1038/s41598-025-89096-9

Radiologie (Heidelb). 2025 Feb 11. doi: 10.1007/s00117-025-01422-4. Online ahead of print.

ABSTRACT

BACKGROUND: The rapid development of large language models (LLMs) opens up new possibilities for the automated processing of medical texts. Transforming unstructured radiology reports into structured data is crucial for efficient use in clinical decision support systems, research, and improving patient care.

OBJECTIVES: What are the challenges of transforming natural language radiology reports into structured data using LLMs? Which methods and architectures are promising? How can the quality and reliability of the extracted data be ensured?

MATERIALS AND METHODS: This article examines current research on the application of LLMs in radiological information processing. Various approaches such as rule-based systems, machine learning, and deep learning models, particularly neural network architectures, are analyzed and compared. The focus is on extracting information such as diagnoses, anatomical locations, findings, and measurements.

RESULTS AND CONCLUSION: LLMs show great potential in transforming reports into structured data. In particular, deep learning models trained on large datasets achieve high accuracies. However, challenges remain, such as dealing with ambiguities, abbreviations, and the variability of linguistic expressions. Combining LLMs with domain-specific knowledge, for example, in the form of ontologies, can further improve the performance of the systems. Integrating contextual information and developing robust evaluation metrics are also important research directions.

PMID:39934245 | DOI:10.1007/s00117-025-01422-4

Sci Rep. 2025 Feb 12;15(1):5157. doi: 10.1038/s41598-024-84922-y.

ABSTRACT

Early detection of ocular diseases is vital to preventing severe complications, yet it remains challenging due to the need for skilled specialists, complex imaging processes, and limited resources. Automated solutions are essential to enhance diagnostic precision and support clinical workflows. This study presents a deep learning-based system for automated classification of ocular diseases using the Ocular Disease Intelligent Recognition (ODIR) dataset. The dataset includes 5,000 patient fundus images labeled into eight categories of ocular diseases. Initial experiments utilized transfer learning models such as DenseNet201, EfficientNetB3, and InceptionResNetV2. To optimize computational efficiency, a novel two-level feature selection framework combining Linear Discriminant Analysis (LDA) and advanced neural network classifiers-Deep Neural Networks (DNN), Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM)-was introduced. Among the tested approaches, the "Combined Data" strategy utilizing features from all three models achieved the best results, with the BiLSTM classifier attaining 100% accuracy, precision, and recall on the training set, and over 98% performance on the validation set. The LDA-based framework significantly reduced computational complexity while enhancing classification accuracy. The proposed system demonstrates a scalable, efficient solution for ocular disease detection, offering robust support for clinical decision-making. By bridging the gap between clinical demands and technological capabilities, it has the potential to alleviate the workload of ophthalmologists, particularly in resource-constrained settings, and improve patient outcomes globally.

PMID:39934192 | DOI:10.1038/s41598-024-84922-y