Do commercial artificial intelligence (AI)-based embryo assessment algorithms differ in their ability to stratify the risk of embryo ploidy, or do they converge towards a similar level of discrimination despite being developed for different clinical objectives? Retrospective cohort study conducted at a single IVF centre, including 1000 blastocysts from 273 patients treated between 2022 and 2023. Embryos were scored using six anonymized commercial AI models (three each of static- or video-based). Ploidy status was determined by trophectoderm biopsy followed by preimplantation genetic testing for aneuploidy (PGT-A). Performance was evaluated using multivariable logistic regression and receiver operating characteristic curves (continuous scores). Euploidy enrichment across within-algorithm score strata and rank concordance (Kendall's τ-b) were assessed, with subgroup analyses by Asociación para el Estudio de la Biología de la Reproducción morphological grade and oocyte age. All six algorithms showed moderate discrimination between euploid and aneuploid embryos (area under the curve range 0.713-0.744, overlapping 95% CI), and their continuous scores were significantly associated with ploidy status (all P < 0.001). For each model, higher score strata were associated with progressive enrichment of euploid embryos, despite substantial overlap in score distributions between ploidy groups. Ranking concordance between algorithms was moderate, indicating that embryos prioritized by different systems frequently differed. These patterns were broadly consistent across morphological grades and maternal age strata. Across 1000 PGT-A-validated blastocysts, commercial AI embryo assessment algorithms converge towards a shared ceiling of ploidy risk stratification performance. These tools should therefore be interpreted as relative risk stratification systems rather than diagnostic tests. While AI scores may support embryo prioritization when PGT-A cannot be performed, limited concordance across algorithms indicates that the choice of algorithm can influence which embryos are prioritized.
Quantum algorithms for selecting a subspace of Hamiltonians to diagonalize have emerged as a promising alternative to variational algorithms in the NISQ era. So far, such algorithms, which include the quantum selected configuration interaction (QSCI) and sample-based quantum diagonalization (SQD), have been formulated in second quantization within Fock space, which leads to inefficient usage of qubit resources. We introduce the first QSCI algorithm developed in the CI-matrix (CIM) framework, which is known to have optimal qubit scaling of exactly⌈log2(N)⌉, where N is the size of the CIM. In addition, we introduce a novel single-bit flip error mitigation which comes at the overhead of a single qubit and we combine this with a stochastic approximate Trotterization evolution adapted from qDRIFT. Simulating benchmark N2 and naphthalene molecules on quantum hardware, our results achieved similar accuracy as SQD methods but with significantly less quantum resources. However, our CIM-QSCI algorithm and SQD methods could not match the performance of classical heat-bath CI (HCI) for the same task. Hence, we introduce an augmented version of QSCI called quantum selected heat-bath CI (QSHCI). This variant replaces classical heat-bath sampling with quantum sampling from QSCI to achieve performance comparable to HCI. We note that a current drawback of our approach is the preprocessing cost of O(N2log⁡N)for constructing the CIM and performing the Pauli decomposition. This can be further improved by considering efficient CIM access models for the stochastic Trotter evolution.
Atherosclerotic disease detection requires models that are both accurate and clinically interpretable. This study proposes a robust and interpretable ensemble learning framework to address this need. We introduce the Progressive Saturated Exponential Loss Adaptive Boosting (ProSEL-Boost) algorithm, which enhances robustness against noisy clinical data through a dynamic loss function with a saturation mechanism. Furthermore, we develop a medical knowledge-guided framework to generate biologically meaningful interaction features based on established pathophysiological pathways, alongside a multi-strategy ensemble method for feature selection. When evaluated on a public coronary artery disease dataset (n = 303) and a proprietary atherosclerosis dataset (n = 284; 37 cases, 247 controls), ProSEL-Boost achieved superior and balanced performance (accuracy = 0.9508, AUC = 0.8807). Crucially, our interpretable feature pipeline identified the glucose-lipid interaction index (TG_FBG_Index) as a potent biomarker (Cohen's d = 1.97, AUC = 0.86), highlighting a synergistic metabolic risk factor. Given the modest sample size, particularly the limited number of positive cases, this study should be viewed primarily as a rigorous methodological validation and hypothesis-generating investigation. The framework provides a compelling rationale for, but does not replace, validation in larger prospective cohorts, offering a transparent methodology for improving early disease detection and generating actionable insights for risk stratification.
To evaluate the impact of image reconstruction algorithms on histogram analysis of iodine concentration (IC) derived from dual-energy CT (DECT) to assess response to first-line chemotherapy in patients with pancreatic ductal adenocarcinoma (PDAC). We retrospectively analyzed 41 patients with PDAC who underwent pancreatic protocol DECT during first-line chemotherapy between January 2021 and January 2024. Iodine-based material decomposition images at the pancreatic phase were reconstructed using hybrid-iterative reconstruction (Hybrid-IR) and deep-learning image reconstruction at medium- and high-intensity levels (DLIR-M and DLIR-H). A region of interest was placed on PDAC, and histogram parameters of tumor IC were extracted from all three reconstructed image sets. These parameters were compared between the response (complete response [CR], partial response [PR], and stable disease [SD]) and non-response (progressive disease [PD]) groups. Receiver-operating-characteristic (ROC) curve analysis was conducted to evaluate the diagnostic performance of significant histogram parameters for differentiating the response and non-response groups. The response and non-response groups were found to differ significantly in standard deviation, energy, and entropy of the Hybrid-IR (P < .001 for all); standard deviation (P = .002), energy (P < .001), and entropy (P < .001) of the DLIR-M; and standard deviation (P = .003), energy (P < .001), entropy (P < .001), and kurtosis (P = .01) of the DLIR-H. Among these 10 parameters, the entropy of the Hybrid-IR (area under the ROC curve, 0.94) and DLIR-M (0.91) demonstrated high and comparable diagnostic performance for differentiating the two groups, with no statistical difference (P = .20), while both outperformed DLIR-H (0.85) (P = .02-.04). The entropy of IC reconstructed with either Hybrid-IR or DLIR-M may serve as an imaging biomarker for assessing chemotherapy response in patients with PDAC. In contrast, DLIR-H may reduce diagnostically relevant texture information and should be used with caution for histogram-based tumor assessment.
This study aims to evaluate cognitive function in patients with Cerebral Small Vessel Disease (CSVD) and investigate its association with variables such as serum Insulin-like Growth Factor-1 (IGF-1). Artificial intelligence algorithms, specifically eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAP), were utilized for analysis and interpretation. A total of 216 patients diagnosed with CSVD were enrolled from the Department of Neurology, Third Affiliated Hospital of Soochow University, between November 2019 and August 2020. Clinical and biochemical data-including triglycerides, total cholesterol, low-density lipoprotein cholesterol, fasting blood glucose, glycosylated hemoglobin (HbA1c), fasting insulin, C-peptide, anti-human insulin antibodies, and IGF-1-were obtained under standardized laboratory protocols. Cognitive function was assessed using the Montreal Cognitive Assessment (MoCA). Based on cognitive performance, patients were categorized into CSVD with cognitive impairment and CSVD without cognitive impairment. The original cohort included 216 patients with CSVD, comprising 50 patients with MCI and 166 patients without MCI. To address class imbalance during model development, SMOTE-NC was applied within the development dataset. The XGBoost model achieved a precision of 0.790, recall of 0.901, F1 score of 0.820, accuracy of 0.833, and Cohen's kappa coefficient of 0.667 in the validation cohort. Feature importance analysis identified key predictors, while SHAP values enabled intuitive visualization of each feature's impact. Decision Curve Analysis (DCA) confirmed the model's clinical utility and net benefit, underscoring its potential for early MCI detection and targeted intervention to improve patient outcomes. Combining XGBoost and SHAP enhances model interpretability, facilitating the identification of critical risk factors such as reduced IGF-1 levels in MoCA-defined MCI in patients with CSVD. This AI-driven approach offers a valuable tool for informing treatment decisions and optimizing healthcare resource allocation.
Industrial robots are a key component of intelligent manufacturing because they improve productivity, precision, and operational reliability. However, long-term operation inevitably introduces wear and other error sources that reduce absolute positioning accuracy and limit precision tasks. To address this issue, this paper develops a two-stage calibrator that combines the advanced social memory optimization algorithm with a neural network optimized by a gradient-based particle swarm optimization scheme, denoted ASMO-GPSONN. In the proposed framework, ASMO identifies robot kinematic errors through memory-guided global exploration, whereas GPSONN compensates the remaining nonlinear residual errors through gradient-corrected swarm refinement. Experiments on two robot calibration datasets, including a real ABB IRB1100 robot, show that the proposed method achieves the best overall calibration accuracy among the compared algorithms. On the held-out test sets, ASMO-GPSONN attains RMSE values of 0.43 mm on D1 and 0.47 mm on D2, demonstrating its practical effectiveness for robot calibration.
Hyponatremia is a prevalent electrolyte disorder with a broad differential diagnosis, which can pose a challenge for determining the underlying etiology. Academic resources teach two diagnostic frameworks designed to guide a physician to the underlying etiology. One approach uses volume status and serum osmolality as earlier branch points, with later consideration of urine osmolality and urine sodium for diagnosis within certain pathways in the branching logic. The other uses serum osmolality, urine osmolality, and spot urine sodium to guide the entire diagnostic reasoning process. These two approaches can ultimately guide clinicians to different diagnoses for the same patient. In this case series, we present an adapted physiology-based (European) algorithm and describe three patients who illustrate key shortcomings of the volume status-based approach to identifying the etiology of hyponatremia. The first case emphasizes that understanding the relationship between osmolality and tonicity adds important nuance to the interpretation of measured serum osmolality. The second case demonstrates how the inherent limitations of the physical examination for volume status contribute to imprecision and inter-individual variation in volume status assessment and underscores the importance of interpreting laboratory values with nuance rather than relying rigidly on strict cutoff values as branch points in clinical reasoning. The third case illustrates that more complex presentations may not fit neatly within any single algorithm and require more nuanced clinical judgment. Together, these cases demonstrate how a diagnostic approach to hyponatremia rooted in physiology provides learners with greater insight into the underlying pathophysiology and may improve diagnostic accuracy.
One of the promises of artificial intelligence (AI) is that it can automate many roles currently performed by humans. However, in the ethics of AI, increased attention is given to the vast amount of labour required for the development, processing, annotation, and implementation of AI. This hidden employment (or 'ghost work') is invisible because of the precarious nature of these jobs, poor working conditions, and the social and economic burdens on employees. For example, to ensure AI does not create undesirable content and that social media platforms do not host it, human data annotators and content moderators (violent content data workers-VCDWs) must read, watch, and listen to content containing torture, rape, bestiality, child sexual abuse, and murder. Through phenomenological analysis, this paper demonstrates that this work has deeply traumatic effects on the affectivity, embodiment, and intersubjectivity of VCDWs. While several other professions (e.g., law enforcement officers, war journalists, and medical professionals) must view violent content, the duration, velocity, and accumulation of exposure to violent images in VCDWs is unprecedented. The outcome of this is trauma from being unable to decipher the ontological status of the content, compassion fatigue due to helplessness to help those suffering, and a myriad of physical, psychological, and interrelation problems for VCDWs. This paper exposes a profession that is systematically concealed, buried under NDAs, outsourced to the Global South, and laundered through ethics-washing and 'responsible AI' branding.
Accurate and timely flood forecasting is essential for issuing effective early warnings and reducing casualties as well as economic losses. However, urban flood forecasting models often struggle to balance computational efficiency with nonlinear representation capability. Forecast errors are rapidly amplified during highly dynamic flooding processes. Assimilating observational data into the model can correct forecast trajectories and reduce overall uncertainties. Existing mainstream data assimilation techniques, such as the ensemble Kalman filter and particle filter, can hardly balance the requirements of nonlinearity and timeliness in urban flood forecasting. Furthermore, effectively utilizing limited observed data to balance the assimilation update frequency and forecast accuracy is critical. To address these challenges, a dynamic urban flood forecasting model coupling the ensemble particle filter with the gradient boosting decision tree was proposed in this study. The performance of state-only and joint state-hyperparameter assimilation for flood prediction at typical ponding points in the central urban area of Zhengzhou was compared. The results indicate that data assimilation improves flood forecast accuracy. State-only assimilation reduces the root mean square error from 0.051-0.084 m to 0.009-0.031 m, while increasing the mean Kling-Gupta efficiency from 0.794 to 0.961. The underlying mechanism whereby incorporating hyperparameters into state variables fails to achieve further performance improvement is analyzed. When the assimilation update frequency is 50 min, an optimal balance between forecast accuracy and observational cost is achieved. Using the sliding window mechanism to identify effective forecast windows for different ponding points provides important support for urban flood emergency decision-making.
This study provides a comprehensive analysis of the application of artificial intelligence (AI) in diagnosing acute traumatic wrist joint injuries (WJIs), including fractures and ligament damage. AI has demonstrated significant potential in identifying fractures and ligament damage. The study highlights the use of various AI technologies and algorithms, including Convolutional Neural Networks (CNNs), Gradient Class Activation Mapping (Grad-CAM), deep learning models, object detection models, automated assessment algorithms, traditional machine learning techniques, data augmentation and preprocessing, Natural Language Processing (NLP), and integration with other imaging modalities. Compared with traditional diagnostic methods, AI offers substantial benefits, such as efficient processing of large datasets, minimizing diagnostic errors and missed cases, aiding in interpreting complex fracture patterns, optimizing workflow, enhancing diagnostic efficiency, and providing comprehensive diagnoses through multimodal integration. AI has significantly improved the precision and efficiency of fracture detection, reduced unnecessary imaging procedures, expedited the diagnostic and reporting process, optimized resource allocation, and improved patient outcomes. However, clinical application of AI faces challenges, including ethical considerations, regulatory hurdles, data privacy and security issues, algorithm transparency and interpretability problems, and unclear liability definitions. The future of AI in diagnosing WJIs is promising, with potential advancements in fracture detection, treatment planning, and rehabilitation strategies. However, challenges such as model validation and training of healthcare professionals must be addressed to fully integrate AI into orthopedic practice and advance the management of WJIs.
Accurate dose deposition in superficial radiotherapy remains challenging due to the inherent skin-sparing effect of megavoltage photon beams and difficulties in ensuring bolus conformity. This study aims to comprehensively evaluate the dosimetric characteristics of the room-temperature-malleable high-density bolus, to evaluate the accuracy of dose calculation and to assess clinical outcomes in a clinical cohort of 55 patients. The dosimetric validation was performed through the use of dose measurements in a solid water phantom on top of which a 1.0 cm thick high-densitybolus was positioned. Absolute depth dose measurements beyond the bolus were obtained by combining an absolute point dose (ion chamber) measurement at 1.0 cm depth beyond the bolus with relative depth dose measurements by using a microDiamond detector. Measurements were performed using 6 megavoltage (MV) flattening-filter-free (FFF) photon beams on an Halcyon and on a TrueBeam treatment unit. Obtained data were compared to dose calculations with both photon dose calculations available in the Eclipse treatment planning system (TPS): Analytical Anisotropic Algorithm (AAA) and Acuros XB (AXB). The impact of non-conforming interfaces was assessed by introducing 0.5 to 2.0 cm thick air cavities between bolus and skin. Skin dose loss due to the air cavities was measured for static open fields and for dynamic sweeping gap fields. Clinical outcomes of 55 patients treated between October 2023 and March 2025 were retrospectively analyzed. Depth dose measurements confirm that a 1.0 cm high-density bolus was sufficient to effectively overcome the build-up region of the 6FFF photon beam. The AXB dose calculation algorithm provided the best agreement between measurements and calculations. Air gaps between the bolus and the skin resulted in clinically relevant skin dose reductions, ranging from approximately 1% to 20%, depending on field size and gap thickness, especially for clinical target volumes < 5.0 cm, large (> 1.0 cm) air gaps, or the use of large static fields. Among the 55 included patients, acute dermatitis toxicity was predominantly Grade 1 (49%) or Grade 2 (47.0%), with a low incidence of severe adverse events (Grade III: 4 %, Grade IV: 0%). The high-density bolus demonstrated satisfying dosimetric reliability, particularly when modeled with the Acuros XB algorithm. The bolus addressed the build-up challenge inherent to MV photons, even in the presence of small air gaps, without increased toxicity in a large cohort of patients. This strategy represents a viable and accessible solution for ensuring accurate and conformal dose delivery in superficial tumors.
The aim of the study was to analyze the influence of various pipelines for preprocessing raw functional magnetic resonance imaging (fMRI) data on the accuracy of classification of subjects into schizophrenia patients and healthy controls using machine learning methods, and to give recommendations for optimizing the data preprocessing pipeline for this task. The study used fMRI data from 72 subjects acquired on a Siemens Magnetom Verio 3T MRI scanner (Siemens Healthineers, Germany) at the National Research Centre "Kurchatov Institute". Seven different preprocessing pipelines were used on each dataset. Feature vectors for classification were constructed from each preprocessed dataset. We applied the following three algorithms for feature vector construction: ReHo, FCM, and FHR. Classification was performed using 15 machine learning methods from the scikit-learn package for each feature set and each preprocessing pipeline. The final accuracy was determined as the maximum accuracy among all machine learning methods. No single optimal preprocessing pipeline for all feature vector construction algorithms was found. Smoothing increased accuracy for the voxel metrics ReHo and FHR, but not for the regional metric FCM. Spatial smoothing was the only preprocessing step that affected classification accuracy for the FHR method. The FHR feature set without smoothing demonstrated accuracy close to chance level. The steps of frequency filtering and median normalization substantially increased the accuracy for both the ReHo and FCM methods, both with and without smoothing. The steps of inhomogeneity correction and slice timing correction did not increase accuracy for the FCM feature set but did improve it by several percent for the ReHo feature set. Application of ICA filtering changed accuracy within a range of 5%, in either a positive or negative direction. Based on the obtained results, the following recommendations can be given for fMRI data preprocessing pipelines for binary classification of subjects into schizophrenia patients and healthy controls using machine learning methods and taking into account the feature vector construction algorithm. The most suitable for the ReHo metric pipeline includes motion correction, normalization, inhomogeneity correction, slice timing correction, frequency filtering, median normalization, and spatial smoothing. The most suitable for the FCM metric pipeline includes motion correction, normalization, inhomogeneity correction, slice timing correction, frequency filtering, and median normalization. Spatial smoothing is the most important factor for the FHR metric. The ICA filtering step does not provide a clear benefit and should therefore be applied with caution.
Artificial intelligence (AI) has revolutionized interventional pulmonology (IP) by enhancing diagnostic accuracy, procedural efficiency, and training standardization. This review synthesizes current advancements and applications of AI across four key domains: education, imaging, navigation, and robotic bronchoscopy systems (RBS). In education, AI-powered convolutional neural networks (CNNs) improve airway structure recognition with an accuracy of 94.7%, thereby reducing learning curves and minimizing diagnostic errors. For imaging, deep learning models achieve lesion identification accuracy comparable to that of senior physicians (area under the curve: 0.940-0.981), enabling real-time intra-procedural guidance. Navigation technologies, such as virtual bronchoscopic navigation (VBN) and electromagnetic navigation bronchoscopy (ENB), enhance accessibility to peripheral lesions, achieving diagnostic yields of 77.9%-94.4% with excellent safety. RBS integrate AI-driven navigation and real-time imaging, achieving biopsy success rates up to 98.8% and diagnostic yield over 90%, while minimizing pneumothorax risks. Despite these advancements, challenges remain in algorithm validation, data privacy, and ethical considerations. AI demonstrates transformative potential in standardizing bronchoscopy practices, expanding access to underserved regions, and advancing personalized treatments. AI significantly enhances the diagnostic accuracy and procedural safety of IP across education, imaging, navigation, and robotic systems; despite ongoing challenges in algorithm validation, AI is propelling the field toward standardized practices and personalized medicine.
Neonatal sepsis (NS) is one of the leading causes of neonatal mortality. The nonspecific clinical manifestations and the limited timeliness of existing biomarkers (such as C-reactive protein) highlight the urgent need for highly accurate diagnostic tools. Neutrophils, as key effector cells of innate immunity, are closely involved in the progression of NS. This study integrated training (GSE69686) and validation (GSE25504) datasets from the GEO database. Neutrophil infiltration characteristics were analyzed utilizing CIBERSORT, and weighted gene co-expression network analysis (WGCNA) was introduced to determine neutrophil-related co-expression modules. Three machine learning algorithms-LASSO, SVM-RFE, and RF-were implemented to cross-screen core diagnostic genes. A combined diagnostic model was distributed based on these genes. NetworkAnalyst was utilized to predict miRNA-TF regulatory networks, and GSVA was conducted to interpret biological functions. Three algorithms identified IL1R2 and METTL7B as core diagnostic genes; the model showed strong reliability. IL1R2 high expression correlated with reduced CD8+ T cells, regulatory T cells, and neutrophils (p<0.05). METTL7B high expression linked positively to B cells and negatively to NK cells/neutrophils. The two genes synergistically cause immune cell dysfunction. Six miRNAs and 15 transcription factors (e.g., NFKB1/RELA, STAT3) regulating these genes were found, involved in inflammation and metabolic reprogramming. Integrating neutrophil infiltration and triple-machine-learning, this study first proposed an IL1R2/METTL7B two-gene panel. The model had high accuracy and generalizability, potentially contributing to NS pathogenesis via immune dysfunction and metabolic reprogramming, supporting rapid diagnostics and targeted interventions.
To compare the performance of different Survival Machine Learning (SML) algorithms in predicting the survival of cancer patients. Data from the Registro Hospitalar de Câncer do Estado de São Paulo (São Paulo State Cancer Registry Hospital) were used, covering the five most incident types of cancer (breast, prostate, lung, colorectal and cervix). Six algorithms were evaluated: Gradient Boosting Survival (GBS), Random Survival Forest (RSF), Support Vector Machine Survival (SVM-Survival), XGBoost Cox, XGBoost Accelerated Failure Time (AFT), and LightGBM. Performance was measured by the Concordance Index (C-Index), C-Index IPCW and Integrated Brier Score (IBS) metrics. The XGBoost AFT model showed the best C-Index results for breast (0.7845), lung (0.7368), colorectal (0.7618), and cervix (0.7726), while GBS was superior for prostate (0.7574). Clinical staging was consistently the most important variable, according to the explainability analysis. The SML algorithms showed good predictive performance, regardless of cancer type, sample size and censoring proportion. These models show potential for subsidizing cancer planning and supporting strategic decisions in the organization of cancer care networks.
Computed tomography (CT) is sometimes necessary during pregnancy to diagnose and treat urgent conditions, but accurate fetal dose estimation remains challenging. Prior studies often relied on uterus dose or simplified fetal models, limiting organ-specific assessment. To develop NCICT3.0 by incorporating detailed computational phantoms of pregnant women and fetuses into NCICT2.0, calculate fetal organ dose coefficients for maternal chest and abdomen-pelvis CT, and evaluate phantom-based estimates against patient-specific CT-based Monte Carlo simulations. Computational phantoms representing fetal ages from 8 to 38 weeks in cephalic presentation were used in Monte Carlo simulations with a validated CT scanner model. Additional breech-presentation models were created for 10, 15, 20, and 25 weeks. Fetal organ dose conversion coefficients (mGy/mGy) were calculated using a general-purpose Monte Carlo code, MCNPX for maternal chest and abdomen-pelvis CT under multiple scan conditions (80, 100, and 120 kVp; head and body filters). Longitudinal tube current modulation (TCM) was modeled using a generic algorithm, and dose changes were assessed relative to size-adjusted fixed-current scans. Phantom-based fetal organ doses were compared with patient-specific CT image-based models simulated using the GPU-based Monte Carlo code MCGPU. Comprehensive fetal organ dose coefficients were generated and implemented into NCICT3.0, demonstrating strong organ- and presentation-dependent variability. In maternal chest CT, fetal doses were low but increased with gestational age, whereas in abdomen-pelvis CT, doses were higher and decreased with age due to increased attenuation. Maternal uterus dose consistently underestimated fetal organ doses. The generic longitudinal TCM algorithm showed a reduction in fetal doses, particularly at early gestation. Comparisons with patient-specific simulations showed that fetal age, maternal abdominal perimeter, and maternal body habitus should be considered when selecting the most appropriate pregnancy phantoms for fetal dose calculations. NCICT3.0 provides a comprehensive fetal organ-level dose library for CT, enabling organ-specific dose estimation and supporting research on dose heterogeneity, risk modeling, and optimization of maternal CT imaging.
Image-domain deep learning denoising may provide a practical post-processing approach for improving low-dose chest CT image quality on scanners without native deep learning reconstruction. This retrospective study evaluated a vendor-independent image-domain denoising algorithm applied to low-dose chest CT on a single 128-slice CT platform and descriptively compared the resulting image-quality metrics with those obtained using a standard-dose iterative reconstruction protocol. This retrospective study included 198 patients who underwent unenhanced chest CT and were assigned to a low-dose CT group (LDCT, n = 99) or a standard-dose CT group (SDCT, n = 99) according to the clinical acquisition protocol. All images were reconstructed using sinogram-affirmed iterative reconstruction (SAFIRE). LDCT images were additionally processed with a vendor-independent image-domain deep learning denoising algorithm (AiR Denoising), generating low-dose AiR-denoised images (LD-AiR). Objective image quality was assessed using attenuation, image noise, signal-to-noise ratio, and contrast-to-noise ratio. Subjective image quality of the lung parenchyma and mediastinal soft tissue was independently evaluated by readers using a 5-point scale. Compared with low-dose SAFIRE (LD-SAFIRE), LD-AiR significantly reduced image noise and increased signal-to-noise ratio and contrast-to-noise ratio across all evaluated regions (all p < 0.05). Subjective image quality scores for both lung parenchyma and mediastinal soft tissue were also significantly higher with LD-AiR than with LD-SAFIRE (all p < 0.05). In the descriptive between-group comparison, the LDCT protocol was associated with an approximately 76% lower effective dose than the SDCT protocol. Most objective and subjective image-quality metrics of LD-AiR did not differ significantly from those of SD-SAFIRE; however, this comparison was based on different patient groups and should not be interpreted as evidence of equivalence or non-inferiority. On the evaluated CT platform, vendor-independent image-domain deep learning denoising improved objective and subjective image quality of low-dose chest CT compared with LD-SAFIRE. These findings suggest that image-domain denoising may have potential utility for image-quality improvement on CT systems without native deep learning reconstruction. Further prospective, within-subject, multicentre studies incorporating lesion-detectability and diagnostic-performance assessment are needed to clarify its clinical applicability.
Projection and backprojection are fundamental operators in tomographic image reconstruction. Existing approaches-namely pixel-driven (PD), ray-driven (RD), and distance-driven (DD)-are subject to well-known trade-offs among interpolation artifacts, geometric flexibility, and computational cost. This study aims to introduce a unified framework that systematically models and resolves the intrinsic limitations of current projection topologies. Within this framework, a novel topology called vector-driven (VD) is proposed to improve geometric flexibility without sacrificing computational efficiency. A three-level hierarchical formulation-comprising topology, algorithm, and implementation-is also developed to isolate the structural origins of artifacts and performance bottlenecks in projection operators. The proposed VD projection and backprojection method represents all tomographic system components as vectors in 3D space and performs non-orthogonal projections that preserve geometric symmetry between forward and backward operations. Experimental validation was conducted using multiple geometric configurations, and quantitative image quality metrics (SSIM, RMSE, PSNR, NCC, UIQI) were compared against PD, RD, and DD baselines. VD produced artifact-free output across all tested geometries and reached near-perfect image-quality scores matching the leading RD-PD baseline combination. VD attained the shortest wall-clock time in both directions, running about 37% faster than an optimised PD algorithm in forward projection and reaching near parity in backprojection, while remaining 31%-57% faster than De Man's DD and 62%-82% faster than Siddon's RD in both directions. This advantage reflects VD's balanced profile of low arithmetic workload of 13.7-24.0 GFLOP per call (giga floating-point operations), high pipeline efficiency with 3.6-4.4 instructions per cycle (IPC), and low DRAM traffic of about 5 GB read per call. VD unifies forward and backward projection within a single vector-based model and, in the configurations tested, performs competitively with established projection topologies in both accuracy and computational cost.
Bone and joint infections (BJIs), including periprosthetic joint infection (PJI), fracture-related infection (FRI), and osteomyelitis, present persistent diagnostic challenges driven by biofilm formation and a high incidence of culture-negative cases. Traditional diagnostic modalities relying on peripheral serum markers and conventional cultures are often limited by insufficient specificity or prolonged turnaround times. This narrative review critically evaluates recent advances in laboratory diagnosis for bone and joint infections, with particular attention to disease-specific applicability across periprosthetic joint infection, fracture-related infection, native vertebral osteomyelitis, diabetic foot osteomyelitis, and other osteomyelitis-related conditions. Current evidence indicates that while traditional serum inflammatory markers are valuable for initial screening, their susceptibility to aseptic inflammatory confounders precludes standalone diagnostic confirmation. In contrast, localized sampling demonstrates significant superiority: novel synovial fluid biomarkers, notably calprotectin and alpha-defensin, accurately reflect the infection microenvironment and offer exceptional diagnostic specificity. At the tissue level, the integration of multiple deep-tissue sampling with preprocessing techniques like sonication has substantially enhanced the recovery of occult biofilm-encased pathogens. Furthermore, targeted and untargeted molecular assays, including multiplex PCR panels, broad-range bacterial PCR, amplicon-based sequencing, and untargeted shotgun metagenomic sequencing, have expanded the diagnostic toolkit for culture-negative, low-virulence, and polymicrobial infections. The diagnostic framework for BJIs has decisively shifted from the pursuit of a solitary "silver bullet" marker toward multimodal, culture-independent assay panels and artificial intelligence-assisted risk stratification algorithms. Future clinical breakthroughs will depend heavily on the global standardization of disease definitions, robust external validation of predictive models, and the seamless integration of advanced laboratory techniques into multidisciplinary team (MDT) workflows.
Early flame detection via Unmanned Aerial Vehicles (UAVs) is crucial for wildfire prevention. However, extreme small scales, non-rigid morphological distortions, and severe background interference in aerial imagery, coupled with limited edge-computing power, pose significant challenges. This study proposes UAV-FlameNet, a lightweight detection algorithm based on YOLOv11. Three innovations are introduced: (1) ADown lossless downsampling to preserve faint flame features; (2) Dynamic Cascaded Flame Aggregation (DCFA) integrating deformable convolutions and variance-aware attention to model thermal radiation and suppress glares; (3) MPDIoU loss to accelerate bounding box convergence. On the LAWD dataset, UAV-FlameNet achieves 83.2% mAP@0.5 with 2.72 M parameters and 5.66 GFLOPs, outperforming YOLOv11 by 2.2% while reducing computational load by 12.1%. The model runs at 121 FPS on GPU and 11.7 FPS on CPU. AV-FlameNet achieves a Pareto-optimal balance among accuracy, robustness, and deployment efficiency, providing a reliable solution for UAV-borne real-time fire early warning systems.