Sci Rep. 2025 Jul 21;15(1):26477. doi: 10.1038/s41598-025-11596-5.

ABSTRACT

One of the challenges in the natural language processing is authorship identification. The proposed research will improve the accuracy and stability of authorship identification by creating a new deep learning framework that combines the features of various types in a self-attentive weighted ensemble framework. Our approach enhances generalization to a great extent by combining a wide range of writing styles representations such as statistical features, TF-IDF vectors, and Word2Vec embeddings. The different sets of features are fed through separate Convolutional Neural Networks (CNN) so that the specific stylistic features can be extracted. More importantly, a self-attention mechanism is presented to smartly combine the results of these specialized CNNs so that the model can dynamically learn the significance of each type of features. The summation of the representation is then passed into a weighted SoftMax classifier with the aim of optimizing performance by taking advantage of the strengths of individual branches of the neural network. The suggested model was intensively tested on two different datasets, Dataset A, which included four authors, and Dataset B, which included thirty authors. Our method performed better than the baseline state-of-the-art methods by at least 3.09% and 4.45% on Dataset A and Dataset B respectively with accuracy of 80.29% and 78.44%, respectively. This self-attention-augmented multi-feature ensemble approach is very effective, with significant gains in state-of-the-art accuracy and robustness metrics of author identification.

PMID:40691694 | DOI:10.1038/s41598-025-11596-5

Sci Rep. 2025 Jul 21;15(1):26446. doi: 10.1038/s41598-025-11780-7.

ABSTRACT

Accurate forecasting of crop yields is essential for ensuring food security and promoting sustainable agricultural practices. Winter wheat, a key staple crop in Pakistan, faces challenges in yield prediction because of the complex interactions among climatic, soil, and environmental factors. This study introduces DeepAgroNet, a novel three-branch deep learning framework that integrates satellite imagery, meteorological data, and soil characteristics to estimate winter wheat yields at the district level in southern Pakistan. The framework employs three leading deep learning models-convolutional neural networks (CNN), recurrent neural networks (RNN), and artificial neural networks (ANN)-trained on detrended yield data from 2017 to 2022. The Google Earth Engine platform was used to process and integrate remote sensing, climate, and soil data. CNN emerged as the most effective model, achieving an R2 value of 0.77 and a forecast accuracy of 98% one month before harvest. The RNN and ANN models also demonstrated moderate predictive capabilities, with R2 values of 0.72 and 0.66, respectively. The results showed that all models achieved less than 10% yield error rates, highlighting their ability to effectively integrate spatial, temporal, and static data. This study emphasizes the importance of deep learning in addressing the limitations of traditional manual methods for yield prediction. By benchmarking the results against Crop Report Services data, this study confirms the reliability and scalability of the proposed framework. The findings demonstrate the potential of DeepAgroNet to improve precision agriculture practices, contributing to food security and sustainable agricultural development in Pakistan. Furthermore, this adaptable framework can serve as a model for similar applications in other agricultural regions around the world.

PMID:40691684 | DOI:10.1038/s41598-025-11780-7

Sci Rep. 2025 Jul 21;15(1):26379. doi: 10.1038/s41598-025-06594-6.

ABSTRACT

Mild to severe anemia is caused by thalassemia, a common genetic disorder affecting over 100 countries worldwide, that results from the abnormality of one or several of the four globin genes. This leads to chronic hemolytic anemia and disrupted synthesis of hemoglobin chains, iron overload, and poor erythropoiesis. Although the diagnosis of thalassemia has improved globally along with the treatment and transfusion support, it is still a major problem in diagnosing in high-prevalence areas like Pakistan. This work aims to assess the performance of numerous combinations of machine learning methods to detect alpha and beta-thalassemia in their minor and major types. These results are obtained from CBC and HPLC analysis. The analyzed models are K-nearest Neighbor (KNN), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost). The study aims to examine the effectiveness of the developed models in discriminating thalassemia variants, especially in the light of Pakistani patients' data. The study found that XGBoost achieved the highest performance on both the CBC and HPLC datasets, with training accuracies of roughly 99.5% for CBC and 99.3% for HPLC. The test accuracy across both datasets was consistently high and thus the best model for detecting thalassemia in this research study. The imported SVM model, slightly less accurate than XGBoost, still has strong performance, particularly on the HPLC data where the cumulative testing accuracy of the model stood at 99.4%. As can be seen from the results, XGBoost specifically shows a very high accuracy of above 99% in the detection of thalassemia types using CBC and HPLC data for Pakistani patients. To the author's knowledge, this research is the first to predict alpha and beta-thalassemia in its major and minor forms using these diagnostic reports. These models indicate that they can offer significant support in detecting thalassemia in resource-constrained settings such as Pakistan. If deep learning is incorporated, even greater accuracy could be achieved.

PMID:40691682 | DOI:10.1038/s41598-025-06594-6

Eur Spine J. 2025 Jul 21. doi: 10.1007/s00586-025-09151-x. Online ahead of print.

ABSTRACT

PURPOSE: Spinal navigation systems require pre- and/or intra-operative 3-D imaging, which expose young patients to harmful radiation. We assessed a scoliosis-specific MRI-protocol that provides T2-weighted MRI and AI-generated synthetic-CT (sCT) scans, through deep learning algorithms. This study aims to compare MRI-based synthetic-CT spinal navigation to CT for safety and accuracy of pedicle screw planning and placement at thoracic and lumbar levels.

METHODS: Spines of 5 cadavers were scanned with thin-slice CT and the scoliosis-specific MRI-protocol (to create sCT). Preoperatively, on both CT and sCT screw trajectories were planned. Subsequently, four spine surgeons performed surface-matched, navigated placement of 2.5 mm k-wires in all pedicles from T3 to L5. Randomization for CT/sCT, surgeon and side was performed (1:1 ratio). On postoperative CT-scans, virtual screws were simulated over k-wires. Maximum angulation, distance between planned and postoperative screw positions and medial breach rate (Gertzbein-Robbins classification) were assessed.

RESULTS: 140 k-wires were inserted, 3 were excluded. There were no pedicle breaches > 2 mm. Of sCT-guided screws, 59 were grade A and 10 grade B. For the CT-guided screws, 47 were grade A and 21 grade B (p = 0.022). Average distance (± SD) between intraoperative and postoperative screw positions was 2.3 ± 1.5 mm in sCT-guided screws, and 2.4 ± 1.8 mm for CT (p = 0.78), average maximum angulation (± SD) was 3.8 ± 2.5° for sCT and 3.9 ± 2.9° for CT (p = 0.75).

CONCLUSION: MRI-based, AI-generated synthetic-CT spinal navigation allows for safe and accurate planning and placement of thoracic and lumbar pedicle screws in a cadaveric model, without significant differences in distance and angulation between planned and postoperative screw positions compared to CT.

PMID:40691585 | DOI:10.1007/s00586-025-09151-x

BMC Med Inform Decis Mak. 2025 Jul 21;25(1):273. doi: 10.1186/s12911-025-03031-6.

ABSTRACT

BACKGROUND: Nosocomial infections are a major cause of morbidity and mortality in the ICU. Earlier identification of these complications may facilitate better clinical management and improve outcomes. We developed a dynamic prediction model that leveraged high-frequency longitudinal data to estimate infection risk 48 h ahead of clinically overt deterioration.

METHODS: We used electronic health record data from consecutive adults who had been treated for > 48 h in a mixed tertiary ICU in the Netherlands enrolled in the Molecular Diagnosis and Risk Stratification of Sepsis (MARS) cohort from 2011 to 2018. All infectious episodes were prospectively adjudicated. ICU-acquired infection (ICU-AI) risk was estimated using a Cox landmark model with high-resolution vital sign data processed via a convolutional neural network (CNN).

RESULTS: We studied 32,178 observation days in 4444 patients and observed 1197 infections, yielding an overall infection risk of 3.5% per ICU day. Discrimination of the composite model was moderate with c-index values varying between 0.64 (95%CI: 0.58-0.69) and 0.72 (95%CI: 0.66-0.78) across timepoints, with some overestimation of ICU-AI risk overall (mean calibration slope 0.58). Compared to 38 common features of infection, a CNN risk score derived from five vital sign signals consistently ranked as a strong predictor of ICU-AI across all time points but did not substantially change risk prediction of ICU-AI.

CONCLUSION: A dynamic modelling approach that incorporates machine learning of high-frequency vital sign data shows promise as a continuous bedside index of infection risk. Further validation is needed to weigh added complexity and interpretability of the deep learning model against potential benefits for clinical decision support in the ICU.

PMID:40691575 | DOI:10.1186/s12911-025-03031-6

BMC Bioinformatics. 2025 Jul 21;26(1):187. doi: 10.1186/s12859-025-06190-5.

ABSTRACT

Bioactive peptides are beneficial or have physiological effects on the life activities of biological organisms. The functions of bioactive peptides are diverse, usually with one or more, so accurately detecting the multiple functions of multi-functional peptides is extremely important. Traditional experimental identification methods are time-consuming, laborious and costly. To overcome these problems, we adopt a computational biology approach and propose a new model BPFun based on deep learning, which can predict seven functions including anticancer, antibacterial, antihypertensive and so on. In BPFun, we obtained the features of bioactive peptides from different aspects, including biological and physicochemical features. Meanwhile, adopting data augmentation to solve the problem of data imbalance. We combine convolutional networks of different scales and Bi-LSTM layers to obtain high-level feature vectors of different features. Finally, the prediction performance is improved by combining these fused features and combining the self-attention mechanism and the Bi-LSTM layer. Our experiments show that BPFun based on five types of sequence features significantly improves the prediction performance of bioactive peptides. Experiments on the test dataset showed that BPFun gets the accuracy and absolute truth value of 0.6577 and 0.6573 on the dataset of seven functional classifications and was superior to other methods. Codes and data are available at https://github.com/291357657/BPFun .

PMID:40691539 | DOI:10.1186/s12859-025-06190-5

BMC Med Imaging. 2025 Jul 21;25(1):292. doi: 10.1186/s12880-025-01831-w.

ABSTRACT

OBJECTIVE: To construct an artificial intelligence (AI)-assisted model for identifying the infraorbital posterior ethmoid cells (IPECs) based on deep learning using sagittal CT images.

METHODS: Sagittal CT images of 277 samples with and 142 samples without IPECs were retrospectively collected. An experienced radiologist engaged in the relevant aspects picked a sagittal CT image that best showed IPECs. The images were randomly assigned to the training and test sets, with 541 sides in the training set and 97 sides in the test set. The training set was used to perform a five-fold cross-validation, and the results of each fold were used to predict the test set. The model was built using nnUNet, and its performance was evaluated using Dice and standard classification metrics.

RESULTS: The model achieved a Dice coefficient of 0.900 in the training set and 0.891 in the additional set. Precision was 0.965 for the training set and 1.000 for the additional set, while sensitivity was 0.981 and 0.967, respectively. A comparison of the diagnostic efficacy between manual outlining by a less-experienced radiologist and AI-assisted outlining showed a significant improvement in detection efficiency (P < 0.05). The AI model aided correctly in identifying and outlining all IPECs, including 12 sides that the radiologist should improve portraying.

CONCLUSION: AI models can help radiologists identify the IPECs, which can further prompt relevant clinical interventions.

PMID:40691526 | DOI:10.1186/s12880-025-01831-w

Nat Mater. 2025 Jul 21. doi: 10.1038/s41563-025-02299-3. Online ahead of print.

NO ABSTRACT

PMID:40691518 | DOI:10.1038/s41563-025-02299-3

Eur Radiol. 2025 Jul 21. doi: 10.1007/s00330-025-11857-x. Online ahead of print.

ABSTRACT

OBJECTIVE: The prevalence of asymptomatic pancreatic cysts is increasing due to advances in imaging techniques. Among these, intraductal papillary mucinous neoplasms (IPMNs) are most common, with potential for malignant transformation, often necessitating close follow-up. This study evaluates novel MRI techniques for the assessment of IPMN.

MATERIALS AND METHODS: From May to December 2023, 59 patients undergoing abdominal MRI were retrospectively enrolled. Examinations were conducted on 3-Tesla scanners using a Deep-Learning Accelerated Half-Fourier Single-Shot Turbo Spin-Echo (HASTEDL) and standard HASTE (HASTES) sequence. Two readers assessed minimum detectable lesion size and lesion-to-parenchyma contrast quantitatively, and qualitative assessments focused on image quality. Statistical analyses included the Wilcoxon signed-rank and chi-squared tests.

RESULTS: HASTEDL demonstrated superior overall image quality (p < 0.001), with higher sharpness and contrast ratings (p < 0.001, p = 0.112). HASTEDL showed enhanced conspicuity of IPMN (p < 0.001) and lymph nodes (p < 0.001), with more frequent visualization of IPMN communication with the pancreatic duct (p < 0.001). Visualization of complex features (dilated pancreatic duct, septa, and mural nodules) was superior in HASTEDL (p < 0.001). The minimum detectable cyst size was significantly smaller for HASTEDL (4.17 mm ± 3.00 vs. 5.51 mm ± 4.75; p < 0.001). Inter-reader agreement was for (к 0.936) for HASTEDL, slightly lower (к 0.885) for HASTES.

CONCLUSION: HASTEDL in IPMN imaging provides superior image quality and significantly reduced scan times. Given the increasing prevalence of IPMN and the ensuing clinical need for fast and precise imaging, HASTEDL improves the availability and quality of patient care.

KEY POINTS: Question Are there advantages of deep-learning-accelerated MRI in imaging and assessing intraductal papillary mucinous neoplasms (IPMN)? Findings Deep-Learning Accelerated Half-Fourier Single-Shot Turbo Spin-Echo (HASTEDL) demonstrated superior image quality, improved conspicuity of "worrisome features" and detection of smaller cysts, with significantly reduced scan times. Clinical relevance HASTEDL provides faster, high-quality MRI imaging, enabling improved diagnostic accuracy and timely risk stratification for IPMN, potentially enhancing patient care and addressing the growing clinical demand for efficient imaging of IPMN.

PMID:40691513 | DOI:10.1007/s00330-025-11857-x

Sci Rep. 2025 Jul 21;15(1):26407. doi: 10.1038/s41598-025-99164-9.

ABSTRACT

Essential genes are necessary for the survival or reproduction of a living organism. The prediction and analysis of gene essentiality can advance our understanding of basic life and human diseases, and further boost the development of new drugs. We propose a snapshot ensemble deep neural network method, DeEPsnap, to predict human essential genes. DeEPsnap integrates the features derived from DNA and protein sequence data with the features extracted or learned from four types of functional data: gene ontology, protein complex, protein domain, and protein-protein interaction networks. More than 200 features from these biological data are extracted/learned which are integrated together to train a series of cost-sensitive deep neural networks. The proposed snapshot mechanism enables us to train multiple models without increasing extra training effort and cost. The experimental results of 10-fold cross-validation show that DeEPsnap can accurately predict human gene essentiality with an average AUROC of 96.16%, AUPRC of 93.83%, and accuracy of 92.36%. The comparative experiments show that DeEPsnap outperforms several popular traditional machine learning models and deep learning models, while all those models show promising performance using the features we created for DeEPsnap. We demonstrated that the proposed method, DeEPsnap, is effective for predicting human essential genes.

PMID:40691502 | DOI:10.1038/s41598-025-99164-9

Int J Radiat Biol. 2025 Jul 21:1-6. doi: 10.1080/09553002.2025.2531908. Online ahead of print.

ABSTRACT

PURPOSE: The development of AI-assisted biodosimetry systems brings significant advances in cytogenetic dosimetry. The introduction of deep learning algorithms has improved the accuracy and speed of chromosome detection and classification in input images, addressing the incomplete reproducibility and time-consuming of manual evaluation. An advanced molecular cytogenetic technique, PNA-FISH, has further improved the clarity and reliability of chromosome identification. We have been developing a deep learning algorithm to automate the detection of chromosomal aberrations in PNA-FISH images, resulting in a more efficient approach to dose assessment, particularly in large-scale nuclear disasters.

CONCLUSION: Integrating AI-assisted biodosimetry systems into the cooperative framework among Advanced Radiation Emergency Medical Support Centers in Japan is expected to support dose assessment in the events of nuclear disasters with mass casualties. However, there are still challenges in the integration of the system into our cooperative framework. Temperature management during blood transport is crucial to prevent coagulation and ensure adequate lymphocyte counts. To improve performance of our AI model, it is necessary to standardize experimental procedures for chromosome image preparations among network members and to further train the learning model. The development of secure and convenient data sharing system is also essential to improve the integrated and practical operation of the network and reduce its running costs. Additionally, development of a user-friendly interface is helpful for all the network members to operate the AI model. We will continue to develop web-based applications for AI-based biodosimetry by considering these requirements to enhance the effectiveness of the biodosimetry network of Japan.

PMID:40690716 | DOI:10.1080/09553002.2025.2531908

PLoS One. 2025 Jul 21;20(7):e0327230. doi: 10.1371/journal.pone.0327230. eCollection 2025.

ABSTRACT

Modern sustainable farming demands precise water management techniques, particularly for crops like potatoes that require high-quality irrigation to ensure optimal growth. This study presents a novel hybrid metaheuristic framework that combines Dipper Throated Optimization (DTO), a bio-inspired algorithm modeled on bird foraging behavior, with Polar Rose Search (PRS) to enhance deep learning models in predictive water quality assessment. The proposed approach integrates binary feature selection and metaheuristic optimization into a unified optimization process, effectively balancing exploration and exploitation to handle complex, high-dimensional datasets. We applied this hybrid strategy to a Radial Basis Function Network (RBFN), and validated its performance improvements through extensive experiments, including ANOVA and Wilcoxon tests for both feature selection and optimization phases. The optimized model achieved a classification accuracy of 99.46%, significantly outperforming classical machine learning and unoptimized deep learning models. These results demonstrate the framework's capability to provide accurate, interpretable, and computationally efficient predictions, which can support smart irrigation decision-making in water-limited agricultural environments, thereby contributing to sustainable crop production and resource conservation.

PMID:40690512 | DOI:10.1371/journal.pone.0327230

J Vis Exp. 2025 Jul 3;(221). doi: 10.3791/68227.

ABSTRACT

As a major cause of death worldwide, cardiovascular diseases-especially arrhythmias-require the creation of precise and automated technologies for early diagnosis and detection. To identify arrhythmias from electrocardiogram (ECG) signals, this paper introduces a deep learning-based classification model that focuses on five main heartbeat types: Normal (N), Left Bundle Branch Block (L), Right Bundle Branch Block (R), Atrial Premature Beat (A), and Premature Ventricular Contraction (V). We leverage Lead I signals from several sources, such as the INCART 12-lead, Sudden Cardiac Death Holter, Supraventricular, and MIT-BIH Arrhythmia databases, yielding more than 3.9 million training and 112,575 testing segments. Examples of data preparation include 180 sample, fixed-window segmentation, Min-Max normalization, and class balancing with the Synthetic Minority Over-sampling Technique (SMOTE). The hybrid architecture uses Transformer layers to model temporal dependencies and 1D Convolutional Neural Networks (CNNs) to extract spatial features. The Adam optimizer with dropout and batch normalization for regularization trains the model. The proposed system outperforms the TN4 model and other cutting-edge benchmarks, achieving 99.99% accuracy, precision, and F1-score across all classes. Feature robustness is further improved by applying deep hybrid architectures and convolutional neural networks, which were motivated by earlier studies. The suggested paradigm advances artificial intelligence-driven, individualized digital healthcare and has great promise for scalable, real-time arrhythmia identification.

PMID:40690419 | DOI:10.3791/68227

IEEE Trans Pattern Anal Mach Intell. 2025 Jul 21;PP. doi: 10.1109/TPAMI.2025.3591076. Online ahead of print.

ABSTRACT

The success of deep learning in computer vision over the past decade has hinged on large labeled datasets and strong pretrained models. In data-scarce settings, the quality of these pretrained models becomes crucial for effective transfer learning. Image classification and self-supervised learning have traditionally been the primary methods for pretraining CNNs and transformer-based architectures. Recently, the rise of text-to-image generative models, particularly those using denoising diffusion in a latent space, has introduced a new class of foundational models trained on massive, captioned image datasets. These models' ability to generate realistic images of unseen content suggests they possess a deep understanding of the visual world. In this work, we present Marigold, a family of conditional generative models and a fine-tuning protocol that extracts the knowledge from pretrained latent diffusion models like Stable Diffusion and adapts them for dense image analysis tasks, including monocular depth estimation, surface normals prediction, and intrinsic decomposition. Marigold requires minimal modification of the pre-trained latent diffusion model's architecture, trains with small synthetic datasets on a single GPU over a few days, and demonstrates state-of-the-art zero-shot generalization. Project page: https://marigoldcomputervision.github.io.

PMID:40690349 | DOI:10.1109/TPAMI.2025.3591076

IEEE Trans Neural Syst Rehabil Eng. 2025 Jul 21;PP. doi: 10.1109/TNSRE.2025.3591254. Online ahead of print.

ABSTRACT

OBJECTIVE: A growing amount of deep learning models for motor imagery (MI) decoding from electroencephalogram (EEG) have demonstrated their superiority over traditional machine learning approaches in offline dataset analysis. However, current online MI-based brain-computer interfaces (BCIs) still predominantly adopt machine learning decoders while falling short of high BCI performance. Yet, the generalization and advantages of deep learning-based EEG decoding in realistic BCI systems remain far unclear.

METHODS: We conduct a randomized and cross-session online MI-BCI study with 2D center-out tasks in 15 BCI-naive subjects. A newly proposed deep learning model named interactive frequency convolutional neural network (IFNet) is leveraged and rigorously compared with the prevailing benchmark namely filter-bank common spatial pattern (FBCSP) for online MI decoding.

RESULTS: Through extensive online analysis, the deep learning decoder consistently outperforms the classical counterpart across various performance metrics. In particular, IFNet significantly improves the average online task accuracy by 20% and 27% in two sessions compared with FBCSP, respectively. Moreover, a significant cross-session training effect is observed by the IFNet model (P = 0.017) while not for the controlled method (P = 0.337). Further offline evaluations also demonstrate the superior performance of IFNet over state-of-the-art deep learning models. Moreover, we present unique behavioral and neurophysiological insights underlying online brain-machine interaction.

CONCLUSION: We present one of the first studies about online MI-BCIs using deep learning, achieving substantially enhanced online performance for continuous BCI control.

SIGNIFICANCE: This study suggests the good utility of deep learning in MI-BCIs and has implications for clinical applications such as stroke rehabilitation.

PMID:40690341 | DOI:10.1109/TNSRE.2025.3591254

Methods Mol Biol. 2025 Jul 22. doi: 10.1007/7651_2025_654. Online ahead of print.

ABSTRACT

Since the generation of embryoid bodies from embryonic stem cells (ESCs), three-dimensional differentiation has been used to mimic developmental processes. To what extent do these in vitro cell types reflect the cells generated by the embryo? We used deep learning (DL) to develop an integrated model of early human development leveraging existing single-cell RNA-seq (scRNA-seq) and using scvi-tools to both integrate and classify cell types. This tool can interrogate in vitro cell types and assign them both identity and provide an entropy score for the reliability of this classification. In this protocol we explain how to use state-of-the-art tools and our associated, publicly available DL models for early embryonic development to explore phenotypes and cell types derived in vitro. Our tools represent an important new resource to interrogate stem cell-based embryo models and the fidelity with which they recapitulate development.

PMID:40690128 | DOI:10.1007/7651_2025_654

Forensic Sci Med Pathol. 2025 Jul 21. doi: 10.1007/s12024-025-01045-0. Online ahead of print.

ABSTRACT

A diagnosis of atherosclerotic cardiovascular disease is critical importance in forensic medicine, particularly because severe atherosclerosis is known to be associated with a high risk of sudden death. In South Korea, the assessment of coronary atherosclerosis during autopsy largely depends on the forensic pathologist's visual measurements, which may limit diagnostic accuracy. The objective of this study was to develop a deep learning algorithm for rapid and precise assessment of coronary atherosclerosis and to identify factors influencing the model's prediction of atherosclerosis severity. A total of 3,717 digital photographs were retrospectively extracted from a database of 1,920 forensic autopsies, with one image each selected for the left anterior descending coronary artery and the right coronary artery. The deep learning algorithm developed in this study demonstrated a high level of agreement (0.988, 95% CI: 0.985-0.990) and absolute agreement (0.986, 95% CI: 0.978-0.991) between predicted and ground truth atherosclerosis values on the test set. The model demonstrated strong overall performance on the test set, achieving a weighted F1-score of 0.904. However, the class-wise F1-scores were 0.957 for mild, 0.785 for moderate, and 0.876 for severe grades, indicating that performance was lowest for the moderate grade. Additionally, decomposition, stent implantation, and thrombi did not have a statistically significant impact on coronary atherosclerosis assessment except for calcification. Although enhancing model performance for moderate grades remains a challenge, this study's findings demonstrate the potential of artificial intelligence as a practical tool for assessing coronary atherosclerosis in autopsy photographs.

PMID:40690102 | DOI:10.1007/s12024-025-01045-0

J Mol Med (Berl). 2025 Jul 21. doi: 10.1007/s00109-025-02575-4. Online ahead of print.

ABSTRACT

Spatial transcriptomics (ST) integrates gene expression with spatial location, enabling precise mapping of cellular distributions and interactions within tissues, and is a key tool for understanding tissue structure and function. Single-cell RNA sequencing (scRNA-seq) data enhances spatial transcriptomics by providing accurate cell type deconvolution, yet existing methods still face accuracy challenges. We propose GraphCellNet, a model combining cell type deconvolution and spatial domain identification, featuring the Kolmogorov-Arnold Network layer (KAN) to enhance nonlinear feature representation and contextual integration. This design addresses ambiguous cell boundaries and high heterogeneity, improving analytical precision. Evaluated using metrics like Pearson correlation coefficient (PCC), structural similarity index (SSIM), root mean square error (RMSE), Jensen-Shannon divergence (JSD), and Adjusted Rand Index (ARI), GraphCellNet has been applied to various systems, yielding new insights. In myocardial infarction, it identified spatial regions with high Trem2 expression associated with metabolic gene signatures in the infarcted heart. In Drosophila development, it uncovered TWEEDLE dynamics. In human heart development, it identified cell compositions and spatial organization across stages, deepening understanding of cellular spatial dynamics and informing regenerative medicine. KEY MESSAGES: A novel deep learning architecture that effectively captures cellular composition and spatial organization in tissue samples. An innovative KAN layer design that improves the modeling of nonlinear gene expression relationships while maintaining computational efficiency. A graph-based spatial domain identification method that leverages the spatial relationships of cell type information to enhance domain recognition accuracy. Demonstration of the framework's applicability in various biological applications, providing new insights into tissue organization and development.

PMID:40690004 | DOI:10.1007/s00109-025-02575-4

Eur Spine J. 2025 Jul 21. doi: 10.1007/s00586-025-09167-3. Online ahead of print.

ABSTRACT

INTRODUCTION: Pedicle screw manufacturer identification is crucial for revision surgery planning; however, this information is occasionally unavailable. We developed a deep learning-based algorithm to identify the pedicle screw manufacturer from plain radiographs.

METHODS: We collected anteroposterior (AP) and lateral radiographs from 276 patients who had thoracolumbar spine surgery with pedicle screws from three international manufacturers. The samples were randomly assigned to training sets (178), validation sets (40), and test sets (58). The algorithm incorporated a convolutional neural network (CNN) model to classify the radiograph as AP and lateral, followed by YOLO object detection to locate the pedicle screw. Another CNN classifier model then identified the manufacturer of each pedicle screw in AP and lateral views. The voting scheme determined the final classification. For comparison, two spine surgeons independently evaluated the same test set, and the accuracy was compared.

RESULTS: The mean age of the patients was 59.5 years, with 1,887 pedicle screws included. The algorithm achieved a perfect accuracy of 100% for the AP radiograph, 98.9% for the lateral radiograph, and 100% when both views were considered. By comparison, the spine surgeons achieved 97.1% accuracy. Statistical analysis revealed near-perfect agreement between the algorithm and the surgeons.

CONCLUSION: We have successfully developed an algorithm for pedicle screw manufacturer identification, which demonstrated excellent accuracy and was comparable to experienced spine surgeons.

PMID:40689982 | DOI:10.1007/s00586-025-09167-3

Invest Ophthalmol Vis Sci. 2025 Jul 1;66(9):55. doi: 10.1167/iovs.66.9.55.

ABSTRACT

PURPOSE: To develop a deep learning algorithm capable of accurately classifying macular neovascularization (MNV) subtypes in patients with treatment-naïve exudative neovascular age-related macular degeneration (AMD) using structural optical coherence tomography (OCT) images.

METHODS: In this retrospective cohort study, a total of 193 eyes with treatment-naïve neovascular AMD were included. Each case was classified into MNV subtypes (type 1, 2, or 3) based on structural OCT features. Convolutional neural network (CNN)-based deep learning models were trained using cross-validation to classify MNV subtypes. Preprocessing included homogenization of image data to optimize use of layer information for classification. Performance metrics included sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve (AUC), with and without homogenization.

RESULTS: Homogenized OCT data improved classification performance compared to non-homogenized data for all models. The highest reported sensitivity and specificity for type 1 MNV was 96.7% and 84.9%; for type 2, 100.0% and 85.5%; and, for type 3, 84.9% and 87.9%, respectively. The AUCs for type 1, 2, and 3 MNV were 0.95, 0.97, and 0.91, respectively. Occlusion sensitivity analysis revealed critical regions for classification, highlighting distinct anatomical differences among MNV subtypes.

CONCLUSIONS: The proposed deep learning model demonstrated high accuracy in classifying MNV subtypes on structural OCT, with improved performance following homogenization. This tool could assist clinicians in accurately and efficiently diagnosing MNV subtypes, potentially influencing treatment decisions and patient outcomes in neovascular AMD.

PMID:40689724 | DOI:10.1167/iovs.66.9.55