Device-to-device (D2D) communication is used to frequently gather and exchange information in various domains. Millimeter-wave research has also incorporated D2D networks. The reliability of multiuser communication is more challenging because of the complex nature of wireless channels. In recent years, the supremacy of the D2D mm-wave communication model has been validated using the outage probability. Generally, the outage and minimize energy consumption to increase the robustness of the network coverage in the D2D mm-wave communication system. In this study, an optimization-enabled Deep Learning (DL) model is introduced to minimize the outage probability and energy consumption. Initially, the simulation of D2D communication was performed, and three types of D2D mm-wave communication coverage probability mechanisms, such as coherent, single-cluster approximation, and non-coherent lower bound, were considered. The minimization of the outage probability is performed using Flamingo Elk Herd Optimization (FEHO). Moreover, transit energy consumption is used to minimize the joint coverage probability by optimally devising a specific threshold. Here, a Deep Spiking Neural Network (DSNN) model is used to create a specific threshold for energy minimization. Furthermore, the performance of the FEHO+DSNN was evaluated by comparing it with existing techniques, where the proposed attained superior performance with 39.056 dBm, and 0.0015 for average transmit power and outage probability.

Fuel-driven hybrid deep learning model for forecasting finger millet prices.

This study investigated the effect of axial length (AL) on peripheral retinal vessel density in children and adolescents and assessed whether deep learning can detect early vascular changes in myopia. Non-mydriatic ultra-widefield imaging was used to capture retinal images. Deep learning models based on Nested U-Net and ResNet34 segmented the vasculature, quantified vessel density in 60-30° and 100-60° fields, and classified AL from fundus images. A total of 679 eyes from 396 children and adolescents were analyzed. Participants were categorized into normal (22.96 ± 0.65mm), medium (24.69 ± 0.50mm), and high (27.32 ± 0.80mm) AL groups. Across both 60-30° and 100-60° fields, the temporal retina displayed higher vessel density, while the inferior retina showed lower density. The normal AL group had significantly higher density than the medium AL group (P < 0.05), which surpassed the high AL group (P < 0.05). In the 60-30° temporal region, vessel density decreased from 7.15 ± 1.17% (normal) to 6.70 ± 1.27% (medium) and 6.16 ± 1.82% (high). Deep learning classification achieved an AUC of 0.9651, with Grad-CAM highlighting the inferotemporal vasculature. As AL increases, peripheral vessel density diminishes. This pattern may suggest a potential prioritization of blood flow to the macular region, although longitudinal studies are required to confirm this hypothesis. These findings suggest that deep learning analysis of ultra-widefield images can reveal subclinical vascular changes, offering a potential tool for early detection of high myopia risk.

Clinically applicable deep learning model for segmentation of the mandibular bone and inferior alveolar canal in CBCT cross-sectional images.

Dependable, contemporaneous sensor systems that can relyably operate despite motion artifacts, noise, and limited power resources are necessary for continuous heart health monitoring. We describe a multimodal technique to monitoring cardiovascular data using earable sensor devices in this paper. To enable self-powered detection of heart vibrations (PCG/PPG-derived waveforms), the suggested wearable employs piezoelectric materials to collect biomechanical energy from human motion. By enhancing high-fidelity cardiac waveforms using a Generative Adversarial Network (GAN) while maintaining morphological properties, medical datasets that are sparse or unbalanced may be addressed. A prediction model based on Kalman Filters is used to monitor cardiac states such heart rate variability (HRV), systolic timing intervals (STI), and probable arrhythmic patterns. This allows for real-time state estimates and noise suppression. The introduction of a lightweight consortium blockchain infrastructure guarantees safe multi-node access, data integrity, and provenance. Remote cardiac telemedicine is a good fit for this paradigm because it offers tamper-resistant medical records, patient identity security, decentralized verification, and visible audit trails. The combined GAN-Kalman architecture decreases noise by 30–45%, increases cardiac anomaly classification accuracy by 12–18%, and shortens real-time prediction latency, according to experimental simulations.

Abstract Background At present, the subtype classification based on clinical symptoms has the limitations of strong subjectivity and unclear biological basis. However, the structural Magnetic Resonance Imaging (sMRI) technology has been proven to capture the morphological changes of the brain. It provides potential biological markers for the objective identification of subtypes of schizophrenia. The research proposes to mine the brain morphological features in sMRI data based on deep learning models to achieve automatic identification of schizophrenia subtypes and provide imaging support for the formulation of subtype-specific treatment plans. Methods A total of 426 patients with schizophrenia aged 18 to 55 (218 cases of paranoid type and 208 cases of undifferentiated type) and 210 healthy controls were selected for the study. All subjects underwent 3.0 T MRI scans to obtain high-resolution T1-weighted sMRI images. The research constructs a 3D Convolutional Neural Network (3D-CNN) model integrating the attention mechanism, taking the preprocessed whole-brain sMRI images as the input. The multi-dimensional morphological features such as volume, thickness and density of gray matter, white matter and cerebrospinal fluid in the brain are automatically extracted through convolutional layers and pooling layers. The attention module focuses on strengthening the feature weights of key brain regions related to schizophrenia, such as the prefrontal lobe and hippocampus. Support Vector Machine (SVM) and Random Forest (RF) were introduced as controls. Results The experimental results show that the constructed 3D-CNN model performs the best in the task of identifying subtypes of schizophrenia, with an overall accuracy rate of 86.4% (95% confidence interval: Among them, the recognition accuracy rate of the paranoid type was 88.2% and the recall rate was 85.3%, the recognition accuracy rate of the undifferentiated type was 84.6% and the recall rate was 87.5%, and the F1 score was 86.9%. In the control models, the accuracy rate of the SVM model was 72.3%, that of the RF model was 68.5%, and the accuracy rate of the 3D-CNN model was 14.1 percentage points higher than that of the optimal traditional model (p&amp;lt;.001). The key differentiating features extracted by the model are mainly concentrated in the prefrontal cortex, hippocampus, thalamus and cingulate gyrus. Among them, the volume reduction of the prefrontal cortex is more significant in patients with paranoid type, while patients with undifferentiated type show combined morphological abnormalities of the hippocampus and thalamus. Discussion The research successfully achieved efficient identification of schizophrenia subtypes from sMRI data through a deep learning model, verifying the specific association between brain morphological features and schizophrenia subtypes. The constructed 3D-CNN model significantly outperformed traditional machine learning methods due to its advantages of automatic feature extraction and attention enhancement in key brain regions. This result provides a non-invasive and objective subtype identification tool based on imaging for clinical practice, which helps to reduce the subjectivity of clinical diagnosis. Future research can expand the sample size of multiple centers, incorporate more subtype categories and combine functional MRI data to further enhance the generalization ability and recognition accuracy of the model.

This paper proposes an intelligent system for verifying the authenticity of academic documents and detecting tampering using artificial intelligence methods. The YOLOv8n deep learning model is used for detecting key areas on an image of an academic document, based on an image of the document being verified (e.g., name, roll number, marks, percentage, institutional seal). The text created by Optical Character Recognition (OCR) from the identified areas is checked against two different ways of determining whether there has been any tampering or forgery of the document using logical validation of the data in question. For example, whether the percentage computed matches the total amount of marks received. If any inconsistency exists between the extracted data from the document being verified, as well as from the logical validation, the areas of the document will be marked on the webpage, which shows both visually the document and the results of the verification. A variety of experiments with test cases show that the current system classifies academic documents into three categories (i.e., legitimate, forged and needing to be confirmed) based on the verification results. Thus, the proposed technique has reduced the time and effort required to manually verify a document and can also assist in detecting academic documents that have possibly been tampered with..

Systematic review of machine and deep learning models for unmanned aerial vehicles cyber threat defense

Introduction Respiratory motion management in radiotherapy for lung cancer patients remains a significant challenge, as it directly affects accurate tumor targeting. Furthermore, unaccounted tumor motion during treatment planning and delivery can lead to imaging artifacts and biased dose distributions, which compromises the accuracy of image-guided radiotherapy. This issue places clinicians in a dilemma between expanding treatment margins, which increases radiation exposure to healthy tissue or risking reduced targeting precision. Methods In this work, a hybrid deep learning model composed of dilated convolutional layers, bidirectional long-short term memory layers, and a generative autoencoder module is proposed to jointly model the spatial and temporal characteristics of respiratory motion, while enabling reconstruction of the physiologically coherent respiratory signals. Each architectural component learns complementary motion-related patterns from respiratory signals to support tumor motion prediction. The model performs motion-range classification, captures abnormal breathing patterns across spatial and temporal domains, reconstructs physiologically coherent respiratory cycles, and predicts tumor motion within an algorithmic validation framework. Results Experimental evaluation demonstrates high motion-range classification performance of 98.37%, including low root-mean square error in motion prediction, while maintaining stable performance across long and complex respiratory signals over multiple breathing cycles. Discussion This study focuses on algorithmic feasibility and establishes a computational foundation for future clinically calibrated and dosimetrically validated models. The findings indicate that the proposed approach can support future motion-aware radiotherapy planning strategies by improving motion characterization at the algorithmic level.

Accurate prediction of wheat prices is crucial for market participants and policymakers because volatility in agricultural markets affects food security and economic planning. This study proposes a hybrid deep-learning-based framework for daily wheat price prediction in Türkiye. The approach first applies an autoencoder to detect and remove anomalous price–quality records from a dataset of 38,019 market transactions collected between June 2022 and May 2023. A weighted ensemble combining Linear Regression, Random Forest, Support Vector Regression and an attention-based Long Short-Term Memory network is then trained on quality parameters and market attributes, with data split into training, validation and test sets. On the independent test set the ensemble achieved a coefficient of determination R2 = 0.9942 and a mean absolute error of 0.1646 TL, outperforming the constituent models. SHAP analysis identifies the price–quality ratio as the most influential feature, while the ablation analysis shows that some of the high accuracy derives from price-derived variables’ strong correlation with the target. Cross-validation confirms robustness and generalization. Overall, the framework provides an effective and interpretable tool for wheat price forecasting, though the short data collection period and single-product focus limit generalizability.

BackgroundHeart failure mortality has risen sharply after years of decline, highlighting the limitations of current risk assessment tools in accuracy, complexity, and cost, and the need for improved predictive models. To address this gap, we developed and validated a deep learning model to improve short-term mortality prediction in heart failure patients.MethodsIn this retrospective study, we leveraged the Medical Information Mart for Intensive Care IV database to develop HF-ECGNet, combining an EfficientNet neural network and a Transformer architecture. We also developed a composite model integrating electrocardiogram-based (ECG) predictions and clinical features. We evaluated model performance using the area under the curve (AUC) and other metrics, with gradient-weighted class activation mapping (Grad-CAM) and Shapley additive explanations (SHAP) analyses for interpretability. We conducted comparisons with N-terminal pro-B-type natriuretic peptide and sequential organ failure assessment (SOFA) scores.ResultsWe analysed a total of 104 844 ECGs from 36 222 admissions. HF-ECGNet achieved an AUC of 0.664 for the first ECG during initial admission, improving to 0.721 for the last ECG. Incorporating three-day ECG data further enhanced performance, with AUCs of 0.691 (first admission) and 0.698 (last admission). HF-ECGNet outperformed NT-proBNP and SOFA. A composite model integrating ECG data and clinical features achieved the highest AUC of 0.725. Grad-CAM identified critical ECG patterns, while SHAP analysis highlighted ECG-derived features as the most influential predictors.ConclusionsHF-ECGNet demonstrates potential as a powerful tool for predicting short-term mortality in heart failure patients. Its innovative architecture and integration of clinical data enable more accurate and interpretable risk stratification. Future multi-centre validation is the critical step to fully ascertain its clinical utility and generalisability.

Objective.Spectral computed tomography (CT) data from photon-counting CT (PCCT) enables material decomposition. Mechanistic approaches such as maximum likelihood estimation are noise sensitive. Deep learning alternatives mitigate this issue, but their accuracy remains limited due to lack of incorporation of underlying physics principles and lack of ground truth data. This study aims to develop and validate a physics-informed deep-learning model, trained on validated simulated data, to decompose spectral CT images into density (ρ)and effective atomic number (Zeff) maps.Methods.The training dataset included simulated abdominal PCCT scans from 32 human models with corresponding ground truth. The scans were obtained at two clinical dose levels, four detector energy thresholds, different iodinated contrast agent concentrations and reconstructed using three clinically-used kernels. A generative adversarial network (GAN) was trained with and without a physics-informed regularization loss to estimateρandZeffmaps. Model performance was evaluated on 16 computational phantoms and validated on 6 clinical cases. A reader study was performed on 30 image slices to assess the comparative performance ofρandZeffmaps to multi-rendered virtual monochromatic images (VMIs) for assessing liver lesion conspicuity.Main results.With physics-informed regularization, NRMSE of 1.29% and 0.68%, SSIM of 0.99 and 0.99, and PSNR of 29.8 dB and 29.04 dB were achieved. A maximum RMSE of 5.45% was achieved on clinical data. Reader study results showedρandZeffimages had higher conspicuity scores compared to VMIs (median: 4.52 vs 4.13; 95% CIs: [4.19, 4.52] vs [4.01, 4.31]). The study showed equivalent conspicuity between VMIs and material images within a ±0.5 margin, though the small sample limits generalization.Significance.This study demonstrates the feasibility of material decomposition using a physics-informed GAN model trained on realistic simulated data. The maps provided equivalent conspicuity under a clinically acceptable margin, with a significantly small number of images for interpretation.

Diagnosing cardiac amyloidosis (CA) on echocardiography can be challenging due to the imaging overlap between CA and more prevalent causes of a hypertrophic phenotype. This study sought to (1) evaluate the performance of artificial-intelligence (AI) derived measurements incorporated into the established multiparametric echocardiographic scoring system to detect CA; (2) develop and validate an AI-based deep-learning model for video-based detection of CA on echocardiography. The study population comprised 5776 patients (CA, 2756; controls, 3020). The training data set included patients from the UK National Amyloidosis Center and Taiwan MacKay Memorial Hospital (CA, 2241; controls, 2130). External test data sets were obtained from the US Duke University Health System (CA, 334; LVH controls, 668) and Japan National Cerebral and Cardiovascular Center (CA, 181; LVH controls, 222). The multiparametric echocardiographic score computed using AI-derived measurements achieved an accuracy of 79.5% (sensitivity, 75.4%; specificity, 81.5%) in the United States cohort and 79.7% (sensitivity, 81.6%; specificity, 78.1%) in the Japan cohort. The deep-learning model demonstrated accuracies of 96.2% (sensitivity, 96.8%; specificity, 95.7%) and 95.8% (sensitivity, 97.3%; specificity, 94.3%) in the internal validation and internal test sets, respectively. External validation of the deep-learning model showed accuracies of 87.5% (sensitivity, 86.6%; specificity, 87.9%) in the United States and 88.4% (sensitivity, 92.3%; specificity, 85.3%) in the Japanese cohort. Subgroup analysis demonstrated that the deep-learning model showed robust discrimination of CA from other hypertrophic phenocopies: CA versus hypertension (area under the curve [AUC], 0.92 [95% CI, 0.91-0.94]), CA versus hypertrophic cardiomyopathy (AUC, 0.91 [95% CI, 0.87-0.94]), CA versus aortic stenosis (AUC, 0.93 [95% CI, 0.90-0.95]), CA versus chronic kidney disease (AUC, 0.93 [95% CI, 0.91-0.95]). The deep-learning model was able to classify a greater proportion of patients compared with the AI-derived multiparametric echocardiographic score and achieved superior diagnostic accuracy (AUC, 0.93 [95% CI, 0.91-0.95] versus AUC, 0.88 [95% CI, 0.85-0.90]; P<0.001). Both the multiparametric echocardiographic score computed from AI-derived measurements and the fully automated deep-learning model can accurately identify patients with CA in globally diverse cohorts, with the deep-learning model providing superior performance.

Accurate and reliable estimation of reference crop evapotranspiration (ET0) in the North Henan Plain is crucial for agricultural water resource management, production, and food supply in China. This study aims to evaluate the performance of deep learning (DL) methods in ET0 estimation and assess the applicability of the developed DL model beyond the training domain. This study utilized historical meteorological data from Zhengzhou City, northern Henan, spanning 2010–2024. Meteorological variables were selected through correlation analysis and maximum information coefficient (MIC). A novel DL model—the TCN-Attention model (TA)—was constructed by incorporating a self-attention mechanism into the temporal convolutional network (TCN) model. This model was compared with two classical DL models—Long Short-Term Memory (LSTM) and TCN. Results indicate: (1) Sunshine duration (n), relative humidity (RH), and maximum temperature (Tmax) are the three most significant features influencing summer maize evapotranspiration; (2) prediction accuracy under the same input scenarios: TA model &gt; TCN model &gt; LSTM model; (3) in scenarios where only temperature data is input, the TA model has the highest prediction accuracy, surpassing the H-S empirical method; and (4) for limited meteorological data, the combination of temperature and humidity was found to be most effective, showing good adaptability and accuracy at different time steps (hourly: R2 = 0.982; daily: R2 = 0.975; weekly: R2 = 0.928). This study highlights the potential of the TA model for estimating reference crop evapotranspiration in the northern Henan Plain, which may provide theoretical guidance for crop irrigation management under future climate change.

BackgroundFascioliasis is a neglected infectious disease affecting agricultural communities worldwide, with the Peruvian Andes among the most severely affected regions. Identifying fine-scale environmental risk patterns could support targeted surveillance and control. We aimed to develop predictive models of Fasciola hepatica infection in humans and sheep using drone-derived environmental indices in a rural Andean community.MethodsWe conducted a cross-sectional study in the Huayllapata community, Cusco, Peru. Demographic, socioeconomic, and georeferenced infection data were collected from households and livestock with fascioliasis diagnosed by stool microscopy. High-resolution multispectral and thermal drone surveys were performed in April 2023 to derive environmental, topographic, and climatic indices. Logistic regression, random forest (RF), XGBoost (XGB), and deep learning models were trained using literature-based or principal component analysis (PCA)-based variable selection strategies. Model performance was evaluated using standard and spatial cross validation approaches. Fine-scale probability surface maps were generated across the study area.ResultsHuman fascioliasis prevalence was 21.3% of households, while sheep prevalence reached 80%. Under standard cross validation, RF achieved the best performance for human infection using the literature-based approach (accuracy = 0.89, sensitivity = 0.99, specificity = 0.88), while XGB performed best using the PCA-based approach (accuracy = 0.85, sensitivity = 0.75, specificity = 0.85). For sheep infection, XGB achieved the highest performance (accuracy = 0.93, sensitivity = 0.65, specificity = 0.93) with literature-based variables and RF performed best under the PCA-based approach (accuracy = 0.85, sensitivity = 0.75, specificity = 0.86). Spatial cross-validation reduced accuracy and specificity across models but preserved high sensitivity. Probability maps revealed marked spatial heterogeneity in predicted risk within the community, with shifts in the location and magnitude of risk zones when spatial dependence was accounted for.ConclusionsIn this single Andean community, machine learning models integrating drone-derived environmental, topographic and climatic indices, successfully identified F. hepatica infection occurrence in humans and sheep. RF and XGB showed the most robust performance under spatial cross-validation, supporting the feasibility of UAV-based approaches for localized F. hepatica risk mapping.Graphical abstractSupplementary InformationThe online version contains supplementary material available at 10.1186/s40249-026-01420-1.

Objective This work aimed to collect joint computed tomography (CT) imaging and peripheral blood transcriptome data from patients with rheumatoid arthritis (RA), and construct a deep learning model for the automatic and precise assessment of bone erosion (BE). It was to screen RA-related inflammation genes regulated by rG4s through bioinformatics methods, explore potential associations between BE imaging phenotypes and molecular regulatory features, and provide hypotheses and clues for investigating the post-transcriptional regulatory mechanisms of RA bone destruction. Methods Clinical data, joint CT images, and peripheral blood RNA sequencing data were collected from the RA group (AG, 148 cases) and the healthy control group (BG, 49 cases) at Yancheng Third People’s Hospital. DESeq2 software was used for differential expression analysis of RNA-seq data. Combined with an inflammation core gene set integrated from multiple databases, RA-related inflammation-related Differentially Expressed Genes (irDEGs) were screened. The rG4detector tool was used to predict rG4s structures in target genes. The Metascape database was used for functional enrichment analysis to identify core candidate genes. An optimized U-Net CNN model was constructed based on the PyTorch framework to achieve automatic segmentation and severity quantification of BE in CT images. Multiple metrics were used to evaluate model performance, and the correlation between candidate gene expression levels and imaging scores was analyzed. Results A total of 67 RA-related irDEGs were screened, of which 42 contained potential rG4s structures. The U-Net CNN model performed excellently in BE segmentation, with pixel-level accuracy, Dice Similarity Coefficient (DSC), sensitivity, and specificity on the test set all at high levels. The model’s quantitative score was significantly correlated with the clinical disease activity score (DAS28). Conclusion CT imaging characteristics of BE in RA patients were closely associated with the expression of rG4s-regulated irDEGs. The deep learning model constructed in this study enabled precise quantification of BE, providing an efficient method for the clinical assessment of RA bone erosion. It also offered a new research perspective and candidate targets for understanding the molecular mechanisms of RA bone destruction at the post-transcriptional regulatory level.

To develop an automated classification system for urinary stone composition by integrating smartphone-based microscopic imaging (TIPSCOPE) with the GoogLeNet architecture, with the goal of enabling rapid, accurate, and cost-effective analysis of stone composition. A total of 140 surgically extracted kidney stone samples were collected and classified into four categories: calcium oxalate (66 cases), uric acid (32 cases), carbonate apatite (26 cases), and magnesium ammonium phosphate hexahydrate (16 cases). Microscopic images of the stones were acquired using the TIPSCOPE device paired with a Realme GT5 smartphone, resulting in a dataset of 840 images. The classification model was trained using the Adam optimizer, with 90% of the dataset allocated for training and 10% reserved for testing. The overall accuracy of the system reached 85.7%. Performance metrics for each category were as follows: uric acid stones: F1 = 0.92 (precision = 0.90, recall = 0.95); magnesium ammonium phosphate hexahydrate stones: F1 = 0.95 (precision = 0.90, recall = 1.00); calcium oxalate stones: F1 = 0.86 (precision = 0.85, recall = 0.88); carbonate apatite stones: F1 = 0.69 (precision = 0.77, recall = 0.63). This study successfully developed a kidney stone composition classification system integrating a smartphone-based microscope with a deep learning model, achieving an overall classification accuracy of 85.7%. The system exhibited strong performance in classifying uric acid and magnesium ammonium phosphate hexahydrate stones. With its low cost, efficiency, and portability, this system offers an economical and practical diagnostic solution for resource-limited regions.

Introduction Cerebral hemorrhage presents a major clinical challenge due to its high mortality and complex pathological characteristics. To address the limitations of traditional diagnostic methods, this study proposes HemorrhageNet, a deep learning framework for automatic classification and prognosis prediction of cerebral hemorrhage. Methods HemorrhageNet integrates multimodal data—including CT and MRI imaging, patient demographics, and clinical parameters—through a dual-path architecture comprising an imaging feature extractor and a clinical feature processor. A graphical propagation layer based on attention mechanisms enables the model to highlight critical hemorrhagic regions, while a multi-task optimization scheme jointly learns classification and prognosis objectives. This design ensures accurate, interpretable, and computationally efficient predictions across diverse patient populations. Building upon this architecture, an adaptive prognostic strategy for cerebral hemorrhage prediction is developed to enhance model generalization and clinical alignment. This strategy incorporates dynamic feature selection to identify the most informative patient-specific attributes, a hierarchical decision-making framework that refines predictions through multi-level reasoning, and uncertainty-aware optimization to quantify confidence and flag ambiguous cases for expert review. These components collectively strengthen interpretability, reduce bias from heterogeneous data, and improve reliability in real-world settings. Results and discussion Extensive experiments on benchmark medical datasets demonstrate that the proposed framework surpasses existing state-of-the-art methods in accuracy, robustness, and transparency. The integration of HemorrhageNet with the adaptive prognostic strategy provides a comprehensive, explainable solution for cerebral hemorrhage management and prognosis assessment.

Background The precise and noninvasive diagnosis of preoperative lymph node metastasis (LNM) in prostate cancer (PC) is challenging. Some studies have studied the application of radiomics-based machine learning (ML) for detecting LNM in PC. However, systematic evidence regarding its diagnostic performance is still lacking. Aim Our study aimed to systematically evaluate the accuracy of radiomics-based ML models in diagnosing LNM in PC, offering evidence-based support for the use of ML in clinical decision-making. Methods Cochrane, PubMed, EMBASE, and Web of Science were searched for eligible studies on the diagnostic performance of radiomics-based ML for LNM in PC until June 11, 2025. The risk of bias in the included studies was evaluated via the Radiomics Quality Score (RQS). Meta-analysis of sensitivity (SEN) and specificity (SPC) was performed using a bivariate mixed-effects model. Subgroup analyses were performed in the meta-analysis based on imaging modality and modeling approach. We conducted meta-analysis on the training and validation sets, respectively. Results A total of 22 studies were included, comprising 13 studies on positron emission tomography (PET)/computed tomography (CT)-based radiomics and nine studies on magnetic resonance imaging (MRI)-based radiomics. In the validation sets, models based on PET/CT yielded a pooled SEN of 0.89 (95% confidence interval (CI): 0.75–0.96), SPC of 0.82 (95% CI: 0.63–0.93), and a summary receiver operating characteristic (SROC) of 0.93 (95% CI: 0.77–0.98). Models based on MRI had a SEN of 0.84 (95% CI: 0.78–0.89), SPC of 0.86 (95% CI: 0.71–0.94), and a SROC of 0.90 (95% CI: 0.71–0.97). Radiomics-based ML models yielded a SEN of 0.85 (95% CI: 0.76–0.91), a SPC of 0.77 (95% CI: 0.66–0.86), and an area under the receiver operating characteristic (AUROC) of 0.89 (95% CI: 0.72–0.96). In contrast, deep learning (DL) models based on radiomics demonstrated a higher SEN of 0.88 (95% CI: 0.75–0.95), SPC of 0.97 (95% CI: 0.58–1.00), and a SROC of 0.95 (95% CI: 0.19–1.00). Conclusions Radiomics demonstrates promising diagnostic performance in detecting LNM in PC. DL models show superior accuracy. Nevertheless, given the limited sample sizes, insufficient external validation, and heterogeneity in imaging protocols, future research should incorporate more multi-center images from different regions. Meanwhile, it is necessary to develop standardized imaging and segmentation protocols to improve transparency and reduce heterogeneity, thereby building more widely applicable and high-performance radiomics-based machine learning models to improve the performance of early detection of LNM in PC patients. Systematic Review Registration https://www.crd.york.ac.uk/prospero/ , identifier PROSPERO CRD420251085724.

Immune checkpoint blockade (ICB) benefits only a subset of sarcoma patients. Biomarkers of response and resistance are needed to help guide patient selection. We analyzed peripheral blood and tumor samples from sarcoma patients treated on five ICB-based clinical trials. Baseline peripheral blood mononuclear cells (PBMCs) underwent 11-color flow cytometry to define T cell immunotypes. Baseline tumor tissue underwent RNA sequencing to classify tumors into four tumor microenvironment (TME) subtypes using consensus clustering of 29 functional gene expression signatures. Associations between immune features and clinical outcomes were assessed. A deep-learning model was applied to baseline hematoxylin and eosin (H&E) slides to detect and quantify lymphoid aggregates in patients with available RNA sequencing. Among 178 patients with PBMCs available for analysis, a proliferative (PRO) circulating T cell immunotype was associated with inferior overall survival (OS) compared with LAG- or LAG+ immunotypes. RNA sequencing from 67 tumors identified an immune-enriched/non-fibrotic TME subtype associated with higher response rate, longer progression-free survival, and longer OS compared to immune-enriched/fibrotic, immune-depleted, and fibrotic subtypes. Automated analysis of 48 baseline H&E slides identified lymphoid aggregates in five tumors; four were classified as immune-enriched and two responded to ICB. Sarcoma patients with a PRO circulating T cell immunotype had inferior outcomes to ICB, while those with an immune-enriched/non-fibrotic TME had superior outcomes. Automated analysis of H&E slides showed promise in identifying patients with an immune-enriched TME. These findings support utilization of a multimodal approach toward identifying predictors of response to immunotherapy in sarcoma.