Cervical cancer remains a major global threat to women’s health, and accurate screening, diagnosis, and individualized care increasingly depend on the integration of multimodal clinical information. This review aims to synthesize current convolutional neural network (CNN)-based methods for multimodal analysis in cervical cancer, with particular focus on how CNN-driven multimodal learning integrates imaging, pathological, molecular, and clinical data to improve cervical cancer diagnosis, prognosis prediction, and clinical decision support. We reviewed common cervical cancer-related data modalities, including imaging, pathology, molecular data, and clinical variables, and summarized their characteristics that influence preprocessing strategies and network design. We further examined CNN-based approaches for modality-specific analysis, multimodal fusion strategies, and hybrid architectures integrating CNNs with attention mechanisms, Transformers, or graph models. Existing studies indicate that CNN-driven multimodal learning can improve sensitivity, diagnostic accuracy, and prognostic performance compared with unimodal approaches. Multimodal fusion enables complementary integration of imaging, pathological, molecular, and clinical information, supporting predictive tasks in diagnosis and prognosis. However, current methods still face challenges such as inter-modal correspondence modeling, data imbalance, limited interpretability, missing modalities, cross-center domain shifts, and privacy constraints. CNN-based multimodal analysis shows considerable potential for advancing precision gynecologic oncology. Future directions include Transformer architectures and foundation models, privacy-preserving federated learning, self-supervised pretraining, and personalized clinical decision-support systems to promote clinical translation.
Autism spectrum disorder (ASD) is a neurodevelopmental condition characterised by social, behavioural and communication traits. The diagnosis of ASD is challenging, particularly due to the heterogeneity of symptoms, overlapping clinical features with other neurodevelopmental conditions and the current reliance on subjective behavioural assessments. Our study investigates the effectiveness of features computed from standard deviation (SD) maps of resting-state functional magnetic resonance imaging (rs-fMRI) in discriminating between ASD and typical development (TD). The rs-fMRI data of TD and ASD considered in this study were obtained from the ABIDE-I and ABIDE-II databases. Initially, the images were pre-processed using a standard pipeline. Further, 3D SD maps were generated, and 110 features were computed from the maps. We fed the features to four machine learning models, such as logistic regression (LR), ridge classifier, gradient boosting, and extreme gradient boosting. We performed the grid search to optimize the parameters and evaluated the models with 5-fold nested cross-validation. We achieved an average 5-fold classification accuracy of 70.71
This study evaluated the biomechanical behavior of zygomatic implants (ZIs) placed via intrasinus and parasinus approaches, focusing on three-dimensional bone-to-implant contact (3D BIC) and stress distribution under vertical, oblique and distal cantilever loadings. Three-dimensional maxillary models were reconstructed from cone-beam computed tomography scans of atrophic maxillae. ZIs were positioned using intrasinus and parasinus trajectories. 3D BIC percentages were calculated for cortical and cancellous bone. Finite element analysis (FEA) simulated vertical (150 N), oblique (150 N), and distal cantilever loadings with varying bending arms (3.2–4.4 mm) to assess von Mises stress in the implant and peri-implant bone. The parasinus pathway achieved a mean cortical 3D BIC of 42.88
Fractional flow reserve (FFR) and resting-full-cycle ratio (RFR) are widely used to determine functional stenosis. However, the results of the FFR and RFR may be discordant. This study aims to decompose aortic (Pa) and distal (Pd) coronary pressure signals under non-hyperemic conditions and explore the variety of intrinsic information among the functional stenosis and non-stenosis of the coronary artery. This study enrolled 21 patients with 21 vessels. The Pa and Pd signal were acquired by Pressure Wire X. Meanwhile, FFR and RFR also be assessment. Four models were tested separately according to the definition of the ischemic cutoff of FFR and RFR. The Holo-Hilbert spectrum (HHS) were used to analyze pa, and the pressure difference ( Δ p = Pa − Pd) signals. Spectral power differences were compared using the Wilcoxon rank-sum test. Both qualitative and quantitative analyses showed that the spectral power within the region of interest (ROI) of the HHS was higher in functional-stenosis than in non-stenosis. The ROI demonstrated higher accuracy and F1-score using the FFR threshold compared with the RFR threshold. While the frequency band of frequency modulation was fixed at 0.5 1.5 Hz, most of areas under curve were greater than 0.8. Using leave-one-out validation, an accuracy of 0.81 was obtained for the FFR ≤ 0.80 threshold. The HHS of Δ p showed good diagnostic performance of the decomposed intrinsic signal in predicting functional stenosis, particularly excluding indeterminate stenosis. The results showed that HHS may provide complementary physiological information for functional stenosis assessment.
To develop and validate a robust, interpretable biocybernetic system for quantitative Parkinson’s disease (PD) assessment from voice recordings, translating vocal disruptions into a continuous physiological metric. We propose a dual-stream architecture fusing deep contextual representations with 40 peripheral neuromotor biomarkers via Multi-Modal Cross-Attention Fusion (MM-CAF). The system was trained on Spanish speech using uncertainty-aware multi-task learning with learnable loss weighting, and externally validated on Italian speech without retraining. A strict subject-independent stratified 5-fold cross-validation was employed. On the Spanish cohort (n = 126), the system achieved accuracy 90.5
This study presents a parsimonious and interpretable radiomic framework for the classification of Alzheimer’s Disease (AD) and its clinical stages using hippocampal MRI. The objective is to evaluate whether compact bivariate models can provide reliable diagnostic performance while maintaining clinical transparency. Structural T1-weighted MRI data from 406 subjects (313 healthy controls and 93 AD patients) were obtained from the OASIS-1 dataset. Automated hippocampal segmentation was followed by the extraction of 25 radiomic features per hemisphere in accordance with Image Biomarker Standardisation Initiative (IBSI) guidelines. All possible two-feature combinations were evaluated using Support Vector Machines with radial basis function kernels. Model training employed stratified 8-fold cross-validation with inverse-frequency class weighting, and final performance was assessed on an independent held-out test set. Bivariate models achieved robust discriminative performance. The combination of hippocampal Volume and Surface Area reached an AUC-ROC of 0.887 in the right hemisphere. In early-stage AD (CDR 0.5), texture-based features (Energy and Sum Entropy) in the left hippocampus provided the most balanced early-stage performance (AUC-ROC = 0.872). In mild-to-moderate stages (CDR ≥ 1), morphological features dominated classification, achieving AUC-ROC values above 0.90. A compact radiomic framework based on two hippocampal features provides an effective and interpretable approach for AD classification. The findings indicate a stage-dependent diagnostic pattern in which textural heterogeneity is more sensitive to early disease, whereas macroscopic atrophy becomes the dominant marker in later stages.
Computed tomography (CT)-guided percutaneous thermal ablation is widely used for minimally invasive treatment of liver tumors where needle path planning critically affects treatment efficacy and safety. This study explores the feasibility of deep reinforcement learning (DRL) for automating needle path planning, framing it as a one-shot anatomical target search problem. A DRL framework was developed comprising a cylindrical conformal environment derived from CT-based three-dimensional models of key anatomical structures (skin, bones, liver, tumor, vessels, spleen), a continuous action space for entry parameters (z, α ) , and a clinically informed reward function integrating five anatomical safety constraints and terminal rewards. The proximal policy optimization (PPO) algorithm was adapted to this non-sequential, constraint-rich search space and trained on clinically annotated cases from a public CT dataset, with the remaining cases for validation. Compared with the conventional rapidly-exploring random tree (RRT) method, the proposed approach achieved a 21.8
In high-stakes medical domains like hemodialysis, models must provide reliable prediction intervals (PIs) that quantify uncertainty, not just point predictions. Standard ensembles are often static. This study proposes and evaluates a novel Uncertainty-Aware Dynamic Weighting Ensemble (UADWE) framework to improve the reliability of blood pressure PIs during hemodialysis. The framework dynamically assigns weights to a pool of base models based on their localized performance. Competence is measured by each model’s historical ability to generate high-quality PIs on similar instances. This adaptive mechanism was validated on a real-world clinical dataset from a hemodialysis center using a rigorous repeated grouped k-fold cross-validation protocol to ensure robust evaluation. The proposed UADWE framework achieved a statistically significant improvement in prediction interval quality (p-value <0.001), obtaining an Interval Score of 50.18 with a small but consistent advantage over Simple Averaging (50.50) and a substantially larger margin over Stacking ensembles (scores > 59). Concurrently, the framework achieved an overall Prediction Interval Coverage Probability of 90 ≈ 11.7). This work introduces a novel dynamic ensemble framework with improved prediction interval quality. We demonstrate that the synergy between a dynamic, uncertainty-aware strategy and a well-curated, diverse model pool enhances reliable interval estimation, though external validation is needed before clinical translation.
Patients with unilateral cleft lip nasal deformity(uCLND) frequently suffer olfactory dysfunction associated with their nasal anatomical deformities, but the role conductive mechanisms play is not understood. Comparative analysis was performed in this study to evaluate diffusion of odorant-laden air in the olfactory airspace between uCLND and healthy subjects. Retrospective computational study with individual-specific anatomically realistic three-dimensional models of 8 healthy subjects with normal nasal anatomy and 7 subjects with uCLND. Three-dimensional nasal models for all subjects were created from computed tomography images. Inspiratory simulations using computational fluid dynamics were performed at 30 L/min to simulate turbulent sniffing conditions. Simulated transport of acetaldehyde was subsequently performed and odorant diffusion in the olfactory mucosa was calculated and compared for subjects with uCLND versus healthy control. Median and interquartile range (IQR) values, as well as p-values and effect sizes for the respective comparisons were calculated. Bilaterally, the median concentration of acetaldehyde in olfactory airspace was significantly greater in healthy group (median = 329.4pg/cm2-s; IQR = 95.6pg/cm2-s) compared to patients with uCLND (median = 15.2pg/cm2-s; IQR = 30.8pg/cm2-s), with p = 0.0012 and a large effect size of 0.93. Similarly, fractions of acetaldehyde diffusing in the olfactory mucosa were Healthy: median = 5.83
Early mortality risk stratification in Acute Respiratory Distress Syndrome (ARDS) remains challenging because conventional approaches often fail to capture the complex interplay between temporal physiological dynamics and the progression of pulmonary structural injury. This study proposed a multimodal adaptive-gating framework to improve mortality prediction in this population. This retrospective study included 102 patients with ARDS, integrating clinical time-series data and dual-time-point dynamic radiomic features. We proposed the Adaptive Gating Multimodal Fusion Network (AdaG-Net), a dual-branch architecture comprising: (1) a clinical branch utilizing Bidirectional Long Short-Term Memory (Bi-LSTM) with channel-wise and additive attention mechanisms; and (2) an imaging branch based on dynamic radiomics. Principal Component Analysis (PCA) was used for dimensionality reduction, while Euclidean distance and difference vectors quantified temporal radiomic evolution. An adaptive gating module dynamically fused multimodal representations. Performance was rigorously assessed using a time-ordered expanding window cross-validation strategy. AdaG-Net achieved a mean AUROC of 0.910 ± 0.018, significantly outperforming radiomics-only, clinical-only, and non-gated multimodal baselines. Incorporating dynamic radiomic evolution features improved imaging branch performance by 8.5
Spinal navigation systems have improved the accuracy and safety of lumbar spinal surgeries. However, traditional manual planning of pedicle screw trajectories remains time-consuming and heavily dependent on surgeon experience. This study presents an automatic pedicle screw trajectory planning system based on preoperative computed tomography images. A 3D U-Net is employed to segment the lumbar vertebrae, with a specific focus on isolating a Clinically Constrained Region that encompasses all feasible screw entry areas while explicitly excluding the articular processes. This segmented region provides essential input for subsequent screw trajectory computation, from which a comprehensive screw trajectory database is constructed. Four bone-mineral-density related strategies were introduced and retrieved from the trajectory database. Among them, two trajectories incorporate the clinically constrained region as an additional anatomical filter. Expert evaluation by experienced spine surgeons indicates that the trajectory extracted through this anatomical filtering closely replicates conventional planning and provides the greatest clinical value. By integrating anatomical structure and biomechanical factors, the proposed system offers effective support for spinal navigation in lumbar procedures.
Auditory attention decoding (AAD) from electroencephalography (EEG) remains challenging because complex spatiotemporal neural patterns are difficult to model, and decoding performance often degrades under short decision windows. This study aimed to develop an efficient and interpretable model for improving short-window AAD performance. A common spatial pattern (CSP)-driven shallow temporal-spatial convolutional neural network (TSCNN) was proposed. CSP spatial filtering was first used to enhance spatially discriminative EEG features associated with attention direction. The filtered signals were then processed by a spatiotemporal convolutional architecture composed of temporal and spatial convolution layers to capture local neural dynamics and cross-channel spatial information. Average pooling and fully connected layers were used for classification. The model was evaluated on the KUL and DTU datasets, and sensitivity and ablation analyses were conducted to assess the contributions of key modules. The proposed model outperformed several state-of-the-art methods on both datasets, particularly under short decision-window conditions. Ablation results further showed that the CSP module, temporal convolution, and spatial convolution each made important contributions to spatial discrimination, short-term dynamic modeling, and multi-channel information integration. Spatial filter visualizations revealed prominent activations in auditory-related cortical regions. The CSP-driven shallow TSCNN provides an effective, interpretable, and low-latency solution for auditory attention decoding and offers methodological support for practical AAD applications.
Early identification of pre-cancerous cervical epithelial changes is clinically critical yet often relies on subjective visual assessment. Quantitative, reproducible imaging biomarkers are therefore required to provide an objective assessment and facilitate artificial intelligence (AI) based automated diagnosis in precancerous cervical screening. This study attempts to identify imaging biomarkers to objectively characterize and differentiate structural changes associated with pre-cancerous cells. The optical microscopic images of superficial and precancerous koilocytotic cells are obtained from a publicly available Pap smear image database with pre-annotated polygon boundaries. Initially, the boundary inconsistencies and minor boundary intersections are corrected in the images. Features such as nuclear area, total cell area, cytoplasm area, nucleus-cytoplasm (NC) ratio, and bending energy are computed from the identified boundaries. To further evaluate the potential for automated differentiation, a logistic regression classification model is employed, and its performance is assessed. Results confirm nuclear enlargement, variation in cytoplasmic space, elevated NC ratios in koilocytes compared with superficial cells, and alterations in nuclear bending energy, attributed to nuclear irregularity, cellular enlargement, and a perinuclear halo. These characteristic features arising due to the effects of human papillomavirus are found to be significant in discriminating koilocytotic cells from superficial cells. The logistic regression model achieved an average five-fold cross-validation weighted F1-score of 85.59
Conventional fetal heart rate monitoring during labor has disadvantages, such as restricting maternal movement and the risk of supine hypotension. This study aimed to evaluate the reliability and safety of a flexible wearable fetal–maternal heart rate monitor, and its impact on the healthcare providers’ and maternal satisfaction. A quasi-experimental design based on convenience sampling was employed to recruit 106 women with singleton pregnancies who were hospitalized and gave birth at Beijing Tsinghua Changgung Hospital between June 2022 and June 2023. The participants were monitored simultaneously using a conventional Doppler fetal heart rate monitor (control group) and a flexible, wearable fetal–maternal heart rate monitor (experimental group) during the onset of labor. Fetal heart rate and uterine contraction data were collected from both monitoring systems throughout labor. In addition, questionnaires were administered to evaluate the satisfaction of healthcare providers and participants with the respective devices. A total of 79 paired fetal heart rate measurements and 49 paired uterine contraction frequency and intensity measurements were analyzed. No significant difference was observed between the flexible and conventional monitoring devices for fetal heart rate (142.86 ± 9.40 vs. 143.66 ± 9.79 bpm, P = 0.0826). Bland–Altman analysis demonstrated a mean bias of -0.80 bpm, with 95
Ultrasound transducers are essential diagnostic tools but are often linked to healthcare-associated infections due to inadequate reprocessing. Conventional chemical disinfectants, though effective, may damage probe surfaces, leave toxic residues, and pose occupational health risks. Cold atmospheric plasma (CAP) has emerged as a promising non-thermal sterilization approach capable of inactivating diverse pathogens without chemical residues. To develop and evaluate a portable dielectric barrier discharge (DBD)-based CAP device for effective, safe, and biocompatible disinfection of ultrasound transducers. A DBD-CAP device was designed to operate at 2 kV, 144 mA, and 26.7 kHz. Optical emission spectroscopy characterized reactive species. Antimicrobial activity was assessed against Escherichia coli, Staphylococcus aureus, carbapenem-resistant Acinetobacter baumannii (CRAB), and methicillin-resistant S. aureus (MRSA). Device integrity and cytocompatibility were evaluated through imaging performance tests and L929 fibroblast cytotoxicity assays. CAP exposure achieved approximately 99
This study investigates whether polylactide (PLA)/hydroxyapatite (HA) composites fabricated by industrially scalable melt-processing can achieve a favorable balance between mechanical performance and in vitro bioactivity for bone-graft substitute applications, and identifies the HA concentration thresholds that govern key material transitions. PLA composites containing 2.5, 5, and 10 wt
Additive manufacturing (AM) enables the fabrication of anatomically precise, patient-specific scaffolds for intervertebral disc (IVD) repair. This study aims to establish an integrated workflow combining computer-aided design (CAD), computational fluid dynamics (CFD), and mechanical-rheological characterization for the development of patient-specific annulus fibrosus (AF) scaffolds and to describe the gelatin–alginate bioink formulation for the fabrication of personalized structures for 3D bioprinting applications with this integrated approach. The L5–S1 AF geometry was reconstructed from cadaveric imaging data and converted into STL format for extrusion-based 3D bioprinting. Scaffold permeability and wall shear stress (WSS) were optimized via Computational Fluid Dynamics (CFD) guided pore design. Gelatin–alginate hydrogels (7–9
This review aims to comprehensively summarize the basic principles of piezoelectric biomaterials, their material classifications, and their applications in tissue engineering. It focuses on elucidating the molecular mechanisms by which piezoelectric stimulation regulates cell behavior and evaluates the translational potential of these materials in regenerative medicine. We systematically reviewed the literature on piezoelectric materials in tissue engineering, covering inorganic, organic, and composite piezoelectric materials. This review integrates in vitro and in vivo study results across various regenerative fields, including bone, cartilage, nerve, skin, cardiovascular, and dental tissues. It also summarizes the key molecular signaling pathways involved in piezoelectric stimulation. Piezoelectric materials effectively convert mechanical energy into localized electrical signals, mimicking endogenous bioelectric cues. They promote tissue regeneration by modulating Ca²⁺ influx via mechanosensitive ion channels, activating integrin-FAK signaling, and regulating pathways such as Wnt/β-catenin, TGF-β, and MAPK/ERK. Applications in bone, cartilage, nerve, skin, cardiovascular, and dental regeneration demonstrate broad therapeutic potential. However, challenges remain in material optimization, long-term biosafety, and clinical translation. Piezoelectric biomaterials offer a promising “self-powered” strategy for tissue regeneration by replicating native electromechanical microenvironments. Future advances will depend on the development of intelligent composites, integration with advanced fabrication technologies, and a deeper understanding of cell-material interactions to enable safe and effective clinical translation.
Previous studies in brain tumor detection primarily focused on architectural modifications to improve performance with limited exploration into class-wise comparisons, intersection-over-union (IoU) threshold adjustments, or targeted data augmentation. The purpose of this study was to investigate the effects of the size and IoU threshold on the tumor detection and classification performance in YOLO architectures. A publicly available Figshare dataset that contains slices with the brain tumors was used for training (n = 2,144), validation (n = 616), and testing (n = 304). YOLOv5n/s/m, v8n/s/m, and v11n/s/m models were used for model development. The data were labeled as one of three classes: meningioma, pituitary or glioma. To evaluate the effect of tumor size, we divided the test dataset into small, medium, and large groups. Moreover, we investigated the effect of IoU threshold on the tumor classification performance. The YOLO models’ performance depended on class type. The average precision at IoU 50
Ultrasound contrast agents (UCAs) enable contrast-enhanced imaging via nonlinear microbubble oscillations, but repeated exposure can deplete the local bubble population and bias perfusion measurements. Conversely, in microbubble-assisted therapies, controlled depletion may be desirable. While destruction is known to increase with peak negative pressure and pulse duration and to decrease with transmit frequency, the role of pulse repetition frequency (PRF) under flow remains incompletely quantified. We developed an acoustic monitoring framework to characterize PRF-dependent microbubble destruction in a vessel phantom. The destruction pulses (3 or 10 MHz; 1 or 3 cycles; 0.5–2.5 MPa peak negative pressure) were applied at PRFs of 0.125–16 kHz, while a 25-MHz transducer continuously recorded downstream M-mode data. Depth-integrated time-intensity curves (TICs) and their time integrals (integrated TIC, ITIC) captured the transition from baseline to depletion. ITICs were described using a simple exponential model, yielding a transient depletion rate constant (b) that provides a kinetic dose-rate descriptor, and a stable destruction percentage (DP) that reflects steady-state survival after replenishment. Both b and DP increased monotonically with PRF, with stronger effects at lower destruction-pulse frequency, higher pressure, and longer pulses. However, b did not uniquely predict DP, indicating that replenishment and exposure history shape steady-state depletion. These metrics provide practical guidance for selecting PRF to preserve contrast signals in imaging or to achieve controlled depletion in therapeutic pulse sequences.
