An Acad Bras Cienc. 2025 Mar 31;97(2):e20240623. doi: 10.1590/0001-3765202520240623. eCollection 2025.

ABSTRACT

The detection and re-identification of vehicles in dynamic environments, such as highways monitored by a swarm of drones, presents significant challenges, particularly due to the variability of images captured from different angles and under various conditions. This scenario necessitates the development of suitable methods that integrate appropriate computational techniques, such as convolutional neural networks (CNN) to address the diversity of drone captures and improve accuracy in detection and re-identification. In this paper, a solution for vehicle detection and Re-ID is proposed, combining CNN techniques VGG16, VGG19, ResNet50, InceptionV3 and EfficientNetV2L. YOLOv4 was selected for detection, while the DeepSORT algorithm was chosen for tracking. The proposed solution considers the generalization capabilities of these techniques with varied images from different drones in different positions. Two datasets were employed: the first is a public dataset from Mendeley used for method evaluation, while the second consists of images and data collected by a swarm of drones. In the first experiment, the best performing network was ResNet50, with an average accuracy of 55%. In the second experiment, the highest accuracy CNN was VGG19, with 91% accuracy. Overall, the techniques were able to distinguish vehicles of different models and adapted to the data captured by drones.

PMID:40172334 | DOI:10.1590/0001-3765202520240623

Radiol Artif Intell. 2025 Apr 2:e240628. doi: 10.1148/ryai.240628. Online ahead of print.

ABSTRACT

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To assess the prognostic value of an open-source deep learning-based chest radiographs (CXR) algorithm, CXR-Lung-Risk, for stratifying respiratory disease mortality risk among an Asian health screening population using baseline and follow-up CXRs. Materials and Methods This single-center, retrospective study analyzed CXRs from individuals who underwent health screenings between January 2004 and June 2018. The CXR-Lung-Risk scores from baseline CXRs were externally tested for predicting mortality due to lung disease or lung cancer, using competing risk analysis, with adjustments made for clinical factors. The additional value of these risk scores beyond clinical factors was evaluated using the likelihood ratio test. An exploratory analysis was conducted on the CXR-Lung-Risk trajectory over a three-year follow-up period for individuals in the highest quartile of baseline respiratory disease mortality risk, using a time-series clustering algorithm. Results Among 36,924 individuals (median age, 58 years [interquartile range: 53-62 years]; 22,352 male), 264 individuals (0.7%) died of respiratory illness, over a median follow-up period of 11.0 years (interquartile range: 7.8- 12.7 years). CXR-Lung-Risk predicted respiratory disease mortality (adjusted hazard ratio [HR] per 5 years: 2.01, 95% CI: 1.76-2.39, P < .001), offering a prognostic improvement over clinical factors (P < .001). The trajectory analysis identified a subgroup with a continuous increase in CXR-Lung-Risk, which was associated with poorer outcomes (adjusted HR for respiratory disease mortality: 3.26, 95% CI: 1.20-8.81, P = .02) compared with the subgroup with a continuous decrease in CXR-Lung-Risk. Conclusion The open-source CXR-Lung-Risk model predicted respiratory disease mortality in an Asian cohort, enabling a two-layer risk stratification approach through an exploratory longitudinal analysis of baseline and follow-up CXRs. ©RSNA, 2025.

PMID:40172326 | DOI:10.1148/ryai.240628

Radiol Artif Intell. 2025 Apr 2:e240507. doi: 10.1148/ryai.240507. Online ahead of print.

ABSTRACT

"Just Accepted" papers have undergone full peer review and have been accepted for publication in Radiology: Artificial Intelligence. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content. Purpose To develop an unsupervised deep learning framework for generalizable blood-brain barrier (BBB) leakage detection using dynamic contrast-enhanced (DCE) MRI, without requiring pharmacokinetic (PK) models and arterial input function (AIF) estimation. Materials and Methods This retrospective study included data from patients who underwent DCE MRI between April 2010 and December 2020. An autoencoder-based anomaly detection (AEAD) identified 1D voxel-wise time-series abnormal signals through reconstruction residuals, separating them into residual leakage signals (RLS) and residual vascular signals (RVS). The RLS maps were evaluated and compared with the volume transfer constant (Ktrans) using the structural similarity index (SSIM) and correlation coefficient (r). Generalizability was tested on subsampled data, and IDH status classification performance was assessed using areas under the receiver operating characteristic curves (AUCs). Results A total of 274 patients were included (164 male; mean age 54.23 ± [SD] 14.66 years). RLS showed high structural similarity (SSIM = 0.91 ± 0.02) and correlation (r = 0.56, P < .001) with Ktrans. On subsampled data, RLS maps showed better correlation with RLS values from original data (0.89 versus 0.72, P < .001), higher PSNR (33.09 dB versus 28.94 dB, P < .001), and higher SSIM (0.92 versus 0.87, P < .001) compared with Ktrans maps. RLS maps also outperformed Ktrans maps in predicting IDH mutation status (AUC = 0.87 [95% CI: 0.83-0.91] versus 0.81 [95% CI: 0.76-0.85], P = .02). Conclusion The unsupervised framework effectively detected blood-brain barrier leakage without PK models and AIF. ©RSNA, 2025.

PMID:40172325 | DOI:10.1148/ryai.240507

JASA Express Lett. 2025 Apr 1;5(4):045201. doi: 10.1121/10.0036356.

ABSTRACT

Laser doppler vibrometers (LDVs) facilitate noncontact speech acquisition; however, they are prone to material-dependent spectral distortions and speckle noise, which degrade intelligibility in noisy environments. This study proposes a data augmentation method that incorporates material-specific and impulse noises to simulate LDV-induced distortions. The proposed approach utilizes a gated convolutional neural network with HiFi-GAN to enhance speech intelligibility across various material and low signal-to-noise ratio (SNR) conditions, achieving a short-time objective intelligibility score of 0.76 at 0 dB SNR. These findings provide valuable insights into optimized augmentation and deep-learning techniques for enhancing LDV-based speech recordings in practical applications.

PMID:40172315 | DOI:10.1121/10.0036356

AJR Am J Roentgenol. 2025 Apr 2. doi: 10.2214/AJR.25.33012. Online ahead of print.

NO ABSTRACT

PMID:40172167 | DOI:10.2214/AJR.25.33012

Front Plant Sci. 2025 Mar 18;16:1498913. doi: 10.3389/fpls.2025.1498913. eCollection 2025.

ABSTRACT

The decline of insect biomass, including pollinators, represents a significant ecological challenge, impacting both biodiversity and ecosystems. Effective monitoring of pollinator habitats, especially floral resources, is essential for addressing this issue. This study connects drone and deep learning technologies to their practical application in ecological research. It focuses on simplifying the application of these technologies. Updating an object detection toolbox to TensorFlow (TF) 2 enhanced performance and ensured compatibility with newer software packages, facilitating access to multiple object recognition models - Faster Region-based Convolutional Neural Network (Faster R-CNN), Single-Shot-Detector (SSD), and EfficientDet. The three object detection models were tested on two datasets of UAV images of flower-rich grasslands, to evaluate their application potential in practice. A practical guide for biologists to apply flower recognition to Unmanned Aerial Vehicle (UAV) imagery is also provided. The results showed that Faster RCNN had the best overall performance with a precision of 89.9% and a recall of 89%, followed by EfficientDet, which excelled in recall but at a lower precision. Notably, EfficientDet demonstrated the lowest model complexity, making it a suitable choice for applications requiring a balance between efficiency and detection performance. Challenges remain, such as detecting flowers in dense vegetation and accounting for environmental variability.

PMID:40171479 | PMC:PMC11959073 | DOI:10.3389/fpls.2025.1498913

Cureus. 2025 Mar 1;17(3):e79871. doi: 10.7759/cureus.79871. eCollection 2025 Mar.

ABSTRACT

Artificial intelligence (AI) has rapidly progressed in the last decade and will inevitably become incorporated into trauma and surgical systems. In such settings, surgeons often need to make high-stakes, time-sensitive, and complex decisions with limited or uncertain information. AI has great potential to augment the pre-operative, intra-operative, and post-operative phases of trauma care. Despite the expeditious advancement of AI, many surgeons lack a foundational understanding of AI terminology, its processes, and potential applications in clinical practice. This narrative review aims to educate general surgeons about the basics of AI, highlight its applications in thoraco-abdominal trauma, and discuss the implications of incorporating its use into the Australian health care system. This review found that studies of AI in trauma care have predominantly focused on machine learning and deep learning applied to diagnostics, risk prediction, and decision-making. Other subfields of AI include natural language processing and computer vision. While AI tools have many potential applications in trauma care, current clinical use is limited. Future prospective, locally validated research is required prior to incorporating AI into clinical practice.

PMID:40171361 | PMC:PMC11958818 | DOI:10.7759/cureus.79871

Front Psychiatry. 2025 Mar 18;16:1452557. doi: 10.3389/fpsyt.2025.1452557. eCollection 2025.

ABSTRACT

Neurological disorders (e.g., Alzheimer's disease and Parkinson's disease) and mental disorders (e.g., depression and anxiety), pose huge challenges to global public health. The pathogenesis of these diseases can usually be attributed to many factors, such as genetic, environmental and socioeconomic status, which make the diagnosis and treatment of the diseases difficult. As research on the diseases advances, so does the body of medical data. The accumulation of such data provides unique opportunities for the basic and clinical study of these diseases, but the vast and diverse nature of the data also make it difficult for physicians and researchers to precisely extract the information and utilize it in their work. A powerful tool to extract the necessary knowledge from large amounts of data is knowledge graph (KG). KG, as an organized form of information, has great potential for the study neurological and mental disorders when it is paired with big data and deep learning technologies. In this study, we reviewed the application of KGs in common neurological and mental disorders in recent years. We also discussed the current state of medical knowledge graphs, highlighting the obstacles and constraints that still need to be overcome.

PMID:40171303 | PMC:PMC11958944 | DOI:10.3389/fpsyt.2025.1452557

Front Oncol. 2025 Mar 18;15:1525650. doi: 10.3389/fonc.2025.1525650. eCollection 2025.

ABSTRACT

In recent years, with the rapid advancement of computer science, artificial intelligence has found extensive applications and has been the subject of significant research within the healthcare industry, particularly in areas such as medical imaging, diagnostics, biomedical engineering, and health data analytics. Artificial intelligence has also made considerable inroads in the diagnosis and treatment of thyroid cancer. This study aims to evaluate the progress, current hotspots, and potential future directions of research on artificial intelligence in the field of thyroid cancer through a bibliometric analysis. This study retrieved literature on the application of artificial intelligence in thyroid cancer from 2004 to 2024 from the Web of Science Core Collection (WoSCC) database. A retrospective bibliometric analysis and visualization study of the filtered data were conducted using VOSviewer, CiteSpace, and the Bibliometrix package in R software. A total of 956 articles from 70 countries/regions were included. China had the highest number of publications, with Shanghai Jiao Tong University (China) being the most prolific research institution. The most prolific author was Wei, X. (n=14), while Haugen, B. R. was the most co-cited author (n=297). The Frontiers in Oncology (35 articles, IF=3.5, Q1) was the most frequently publishing journal, and Thyroid (cited 1,705 times) was the most co-cited journal. Keywords such as 'ultrasound,' 'deep learning,' and 'diagnosis' indicate research hotspots in this field. This study provides a comprehensive exposition of the current advancements, emerging trends, and future directions of artificial intelligence in thyroid cancer research. It serves as a valuable resource for clinicians and researchers, offering a systematic understanding of key focal areas in the field, thereby assisting in the identification and determination of future research trajectories.

PMID:40171256 | PMC:PMC11958942 | DOI:10.3389/fonc.2025.1525650

Front Endocrinol (Lausanne). 2025 Mar 18;16:1485311. doi: 10.3389/fendo.2025.1485311. eCollection 2025.

ABSTRACT

OBJECTIVE: To systematically review and meta-analyze the effectiveness of deep learning algorithms applied to optical coherence tomography (OCT) and retinal images for the detection of diabetic retinopathy (DR).

METHODS: We conducted a comprehensive literature search in multiple databases including PubMed, Cochrane library, Web of Science, Embase and IEEE Xplore up to July 2024. Studies that utilized deep learning techniques for the detection of DR using OCT and retinal images were included. Data extraction and quality assessment were performed independently by two reviewers. Meta-analysis was conducted to determine pooled sensitivity, specificity, and diagnostic odds ratios.

RESULTS: A total of 47 studies were included in the systematic review, 10 were meta-analyzed, encompassing a total of 188268 retinal images and OCT scans. The meta-analysis revealed a pooled sensitivity of 1.88 (95% CI: 1.45-2.44) and a pooled specificity of 1.33 (95% CI: 0.97-1.84) for the detection of DR using deep learning models. All of the outcome of deep learning-based optical coherence tomography ORs ≥0.785, indicating that all included studies with artificial intelligence assistance produced good boosting results.

CONCLUSION: Deep learning-based approaches show high accuracy in detecting diabetic retinopathy from OCT and retinal images, supporting their potential as reliable tools in clinical settings. Future research should focus on standardizing datasets, improving model interpretability, and validating performance across diverse populations.

SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/, identifier CRD42024575847.

PMID:40171193 | PMC:PMC11958191 | DOI:10.3389/fendo.2025.1485311

Digit Health. 2025 Mar 31;11:20552076251332999. doi: 10.1177/20552076251332999. eCollection 2025 Jan-Dec.

ABSTRACT

OBJECTIVE: Gait analysis plays a pivotal role in evaluating walking abilities, with recent advancements in digital health stressing the importance of efficient data collection methods. This study aims to classify nine gait types including one normal and eight abnormal gaits, using sequential network-based models and diverse feature combinations obtained from insole sensors.

METHODS: The dataset was collected using insole sensors from subjects performing 15 m walking with designated gait types. The sensors incorporated pressure sensors and inertial measurement units (IMUs), along with the center of pressure engineered from the pressure readings. A number of deep learning architectures were evaluated for their ability to classify the gait types, focusing on feature sets including temporal parameters, statistical features of pressure signals, center of pressure data, and IMU data. Ablation studies were also conducted to assess the impact of combining features from different modalities.

RESULTS: Our results demonstrate that models incorporating IMU features outperform those using different combinations of modalities including individual feature sets, with the top-performing models achieving F1-scores of up to 90% in sample-wise classification and 92% in subject-wise classification. Additionally, an ablation study reveals the importance of considering diverse feature modalities, including temporal parameters, statistical features from pressure signals, center of pressure data, and IMU data, for comprehensive gait classification.

CONCLUSION: Overall, this study successfully developed deep sequential models that effectively classify nine different gait types, with the ablation study underscoring the potential for integrating features from diverse domains to enhance clinical applications, such as intervention for gait-related disorders.

PMID:40171146 | PMC:PMC11960168 | DOI:10.1177/20552076251332999

Front Physiol. 2025 Mar 18;16:1511716. doi: 10.3389/fphys.2025.1511716. eCollection 2025.

ABSTRACT

INTRODUCTION: Lung nodule detection is a crucial task for diagnosis and lung cancer prevention. However, it can be extremely difficult to identify tiny nodules in medical images since pulmonary nodules vary greatly in shape, size, and location. Further, the implemented methods have certain limitations including scalability, robustness, data availability, and false detection rate.

METHODS: To overcome the limitations in the existing techniques, this research proposes the Cnidaria Herd Optimization (CHO) algorithm-enabled Bi-directional Long Short-Term Memory (CHSTM) model for effective lung nodule detection. Furthermore, statistical and texture descriptors extract the significant features that aid in improving the detection accuracy. In addition, the FC2R segmentation model combines the optimized fuzzy C-means clustering algorithm and the Resnet -101 deep learning approach that effectively improves the performance of the model. Specifically, the CHO algorithm is modelled using the combination of the induced movement strategy of krill with the time control mechanism of the cnidaria to find the optimal solution and improve the CHSTM model's performance.

RESULTS: According to the experimental findings of a performance comparison between other established methods, the FC2R + CHSTM model achieves 98.09% sensitivity, 97.71% accuracy, and 97.03% specificity for TP 80 utilizing the LUNA-16 dataset. Utilizing the LIDC/IDRI dataset, the proposed approach attained a high accuracy of 97.59%, sensitivity of 96.77%, and specificity of 98.41% with k-fold validation outperforming the other existing techniques.

CONCLUSION: The proposed FC2R + CHSTM model effectively detects lung nodules with minimum loss and better accuracy.

PMID:40171113 | PMC:PMC11959082 | DOI:10.3389/fphys.2025.1511716

J Appl Crystallogr. 2025 Feb 28;58(Pt 2):513-522. doi: 10.1107/S1600576725000974. eCollection 2025 Apr 1.

ABSTRACT

Recent advancements in X-ray sources and detectors have dramatically increased data generation, leading to a greater demand for automated data processing. This is particularly relevant for real-time grazing-incidence wide-angle X-ray scattering (GIWAXS) experiments which can produce hundreds of thousands of diffraction images in a single day at a synchrotron beamline. Deep learning (DL)-based peak-detection techniques are becoming prominent in this field, but rigorous benchmarking is essential to evaluate their reliability, identify potential problems, explore avenues for improvement and build confidence among researchers for seamless integration into their workflows. However, the systematic evaluation of these techniques has been hampered by the lack of annotated GIWAXS datasets, standardized metrics and baseline models. To address these challenges, we introduce a comprehensive framework comprising an annotated experimental dataset, physics-informed metrics adapted to the GIWAXS geometry and a competitive baseline - a classical, non-DL peak-detection algorithm optimized on our dataset. Furthermore, we apply our framework to benchmark a recent DL solution trained on simulated data and discover its superior performance compared with our baseline. This analysis not only highlights the effectiveness of DL methods for identifying diffraction peaks but also provides insights for further development of these solutions.

PMID:40170972 | PMC:PMC11957406 | DOI:10.1107/S1600576725000974

Stat Biosci. 2025 Apr;17(1):132-150. doi: 10.1007/s12561-023-09394-6. Epub 2023 Oct 28.

ABSTRACT

Estimating treatment effects is of great importance for many biomedical applications with observational data. Particularly, interpretability of the treatment effects is preferable for many biomedical researchers. In this paper, we first provide a theoretical analysis and derive an upper bound for the bias of average treatment effect (ATE) estimation under the strong ignorability assumption. Derived by leveraging appealing properties of the weighted energy distance, our upper bound is tighter than what has been reported in the literature. Motivated by the theoretical analysis, we propose a novel objective function for estimating the ATE that uses the energy distance balancing score and hence does not require the correct specification of the propensity score model. We also leverage recently developed neural additive models to improve interpretability of deep learning models used for potential outcome prediction. We further enhance our proposed model with an energy distance balancing score weighted regularization. The superiority of our proposed model over current state-of-the-art methods is demonstrated in semi-synthetic experiments using two benchmark datasets, namely, IHDP and ACIC, as well as is examined through the study of the effect of smoking on the blood level of cadmium using NHANES.

PMID:40170916 | PMC:PMC11957463 | DOI:10.1007/s12561-023-09394-6

Health Data Sci. 2025 Apr 1;5:0256. doi: 10.34133/hds.0256. eCollection 2025.

ABSTRACT

Background: Multimodal large language models (LLMs) have shown potential in various health-related fields. However, many healthcare studies have raised concerns about the reliability and biases of LLMs in healthcare applications. Methods: To explore the practical application of multimodal LLMs in skin disease identification, and to evaluate sex and age biases, we tested the performance of 2 popular multimodal LLMs, ChatGPT-4 and LLaVA-1.6, across diverse sex and age groups using a subset of a large dermatoscopic dataset containing around 10,000 images and 3 skin diseases (melanoma, melanocytic nevi, and benign keratosis-like lesions). Results: In comparison to 3 deep learning models (VGG16, ResNet50, and Model Derm) based on convolutional neural network (CNN), one vision transformer model (Swin-B), we found that ChatGPT-4 and LLaVA-1.6 demonstrated overall accuracies that were 3% and 23% higher (and F1-scores that were 4% and 34% higher), respectively, than the best performing CNN-based baseline while maintaining accuracies that were 38% and 26% lower (and F1-scores that were 38% and 19% lower), respectively, than Swin-B. Meanwhile, ChatGPT-4 is generally unbiased in identifying these skin diseases across sex and age groups, while LLaVA-1.6 is generally unbiased across age groups, in contrast to Swin-B, which is biased in identifying melanocytic nevi. Conclusions: This study suggests the usefulness and fairness of LLMs in dermatological applications, aiding physicians and practitioners with diagnostic recommendations and patient screening. To further verify and evaluate the reliability and fairness of LLMs in healthcare, experiments using larger and more diverse datasets need to be performed in the future.

PMID:40170800 | PMC:PMC11961048 | DOI:10.34133/hds.0256

MAbs. 2025 Dec;17(1):2483944. doi: 10.1080/19420862.2025.2483944. Epub 2025 Apr 1.

ABSTRACT

Highly concentrated antibody solutions are necessary for developing subcutaneous injections but often exhibit high viscosities, posing challenges in antibody-drug development, manufacturing, and administration. Previous computational models were only limited to a few dozen data points for training, a bottleneck for generalizability. In this study, we measured the viscosity of a panel of 229 monoclonal antibodies (mAbs) to develop predictive models for high concentration mAb screening. We developed DeepViscosity, consisting of 102 ensemble artificial neural network models to classify low-viscosity (≤20 cP) and high-viscosity (>20 cP) mAbs at 150 mg/mL, using 30 features from a sequence-based DeepSP model. Two independent test sets, comprising 16 and 38 mAbs with known experimental viscosity, were used to assess DeepViscosity's generalizability. The model exhibited an accuracy of 87.5% and 89.5% on both test sets, respectively, surpassing other predictive methods. DeepViscosity will facilitate early-stage antibody development to select low-viscosity antibodies for improved manufacturability and formulation properties, critical for subcutaneous drug delivery. The webserver-based application can be freely accessed via https://devpred.onrender.com/DeepViscosity.

PMID:40170162 | DOI:10.1080/19420862.2025.2483944

BMC Bioinformatics. 2025 Apr 1;26(1):96. doi: 10.1186/s12859-025-06092-6.

ABSTRACT

BACKGROUND: Chromatin loops are critical for the three-dimensional organization of the genome and gene regulation. Accurate identification of chromatin loops is essential for understanding the regulatory mechanisms in disease. However, current mainstream detection methods rely primarily on single-source data, such as Hi-C, which limits these methods' ability to capture the diverse features of chromatin loop structures. In contrast, multi-source data integration and deep learning approaches, though not yet widely applied, hold significant potential.

RESULTS: In this study, we developed a method called DconnLoop to integrate Hi-C, ChIP-seq, and ATAC-seq data to predict chromatin loops. This method achieves feature extraction and fusion of multi-source data by integrating residual mechanisms, directional connectivity excitation modules, and interactive feature space decoders. Finally, we apply density estimation and density clustering to the genome-wide prediction results to identify more representative loops. The code is available from https://github.com/kuikui-C/DconnLoop .

CONCLUSIONS: The results demonstrate that DconnLoop outperforms existing methods in both precision and recall. In various experiments, including Aggregate Peak Analysis and peak enrichment comparisons, DconnLoop consistently shows advantages. Extensive ablation studies and validation across different sequencing depths further confirm DconnLoop's robustness and generalizability.

PMID:40170155 | DOI:10.1186/s12859-025-06092-6

BMC Med Inform Decis Mak. 2025 Apr 1;25(1):156. doi: 10.1186/s12911-025-02983-z.

ABSTRACT

BACKGROUND: Risk stratification of COVID-19 patients can support therapeutic decisions, planning and resource allocation in the hospital. In times of high incidence, a prognostic model based on data efficiently retrieved from one source can enable fast decision support.

METHODS: A model was developed to identify patients at risk of developing severe COVID-19 within one month based on their age, sex and imaging features extracted from the thoracic computed tomography (CT). The model was trained on publicly available data from the Study of Thoracic CT in COVID-19 (STOIC) challenge and validated on unseen data from the same study and an external, multicentric dataset. The model, trained on data acquired before any variant of concern dominated, was assessed separately on data collected at later stages of the pandemic when the delta and omicron variants were most prevalent.

RESULTS: A logistic regression based on handcrafted features was found to perform on par with a direct deep learning approach, and the former was selected for simplicity. Volumetric and intensity-based features of lesions and healthy lung parenchyma proved most predictive, in addition to patient age and sex. The model reached an area under the curve of 0.78 on the challenge test set and 0.74 on the external test set. The performance did not drop for the subset acquired at a later stage of the pandemic.

CONCLUSIONS: A logistic regression utilizing features from thoracic CT and its metadata can provide rapid decision support by estimating short-term COVID-19 severity. Its stable performance underscores its potential for real-world clinical integration. By enabling rapid risk stratification using readily available imaging data, this approach can support early clinical decision-making, optimize resource allocation, and improve patient management, particularly during surges in COVID-19 cases. Furthermore, this study provides a foundation for future research on prognostic modelling in respiratory infections.

PMID:40170034 | DOI:10.1186/s12911-025-02983-z

BMC Ophthalmol. 2025 Apr 1;25(1):162. doi: 10.1186/s12886-025-03988-y.

ABSTRACT

PURPOSE: To compare two artificial intelligence (AI) models, residual neural networks ResNet-50 and ResNet-101, for screening thyroid eye disease (TED) using frontal face photographs, and to test these models under clinical conditions.

METHODS: A total of 1601 face photographs were obtained. These photographs were preprocessed by cropping to a region centered around the eyes. For the deep learning process, photographs from 643 TED patients and 643 healthy individuals were used for training the ResNet models. Additionally, 81 photographs of TED patients and 74 of normal subjects were used as the validation dataset. Finally, 80 TED cases and 80 healthy subjects comprised the test dataset. For application tests under clinical conditions, data from 25 TED patients and 25 healthy individuals were utilized to evaluate the non-inferiority of the AI models, with general ophthalmologists and fellowships as the control group.

RESULTS: In the test set verification of the ResNet-50 AI model, the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, and specificity were 0.94, 0.88, 0.64, and 0.92, respectively. For the ResNet-101 AI model, these metrics were 0.93, 0.84, 0.76, and 0.92, respectively. In the application tests under clinical conditions, to evaluate the non-inferiority of the ResNet-50 AI model, the AUC, accuracy, sensitivity, and specificity were 0.82, 0.82, 0.88, and 0.76, respectively. For the ResNet-101 AI model, these metrics were 0.91, 0.84, 0.92, and 0.76, respectively, with no statistically significant differences between the two models for any of the metrics (all p-values > 0.05).

CONCLUSIONS: Face image-based TED screening using ResNet-50 and ResNet-101 AI models shows acceptable accuracy, sensitivity, and specificity for distinguishing TED from healthy subjects.

PMID:40169995 | DOI:10.1186/s12886-025-03988-y

BMC Oral Health. 2025 Apr 1;25(1):465. doi: 10.1186/s12903-025-05827-4.

ABSTRACT

BACKGROUND: The aim was to fully automate molar teeth developmental staging and to comprehensively analyze a wide range of deep learning models' performances for molar tooth germ detection on panoramic radiographs.

METHODS: The dataset consisted of 210 panoramic radiographies. The data were obtained from patients aged between 5 and 25 years. The stages of development of molar teeth were divided into 4 classes such as M1, M2, M3 and M4. 9 different convolutional neural network models, which were Cascade R-CNN, YOLOv3, Hybrid Task Cascade(HTC), DetectorRS, SSD, EfficientNet, NAS-FPN, Deformable DETR and Probabilistic Anchor Assignment(PAA), were used for automatic detection of these classes. Performances were evaluated by mAP for detection localization performance and confusion matrices, giving metrics of accuracy, precision, recall and F1-scores for classification part.

RESULTS: Localization performance of the models varied between 0.70 and 0.86 while average accuracy for all classes was between 0.71 and 0.82. The Deformable DETR model provided the best performance with mAP, accuracy, recall and F1-score as 0.86, 0.82, 0.86 and 0.86 respectively.

CONCLUSIONS: Molar teeth were automatically detected and categorized by modern artificial intelligence techniques. Findings demonstrated that detection and classification ability of deep learning models were promising for molar teeth development staging. Automated systems have a potential to alleviate the burden and assist dentists.

TRIAL REGISTRATION: This is retrospectively registered with the number 2023-1216 by the university ethical committee.

PMID:40169944 | DOI:10.1186/s12903-025-05827-4