Huynh Quang Linh

Traditional conic-based models rely on conic constants, which exhibit high interindividual variability and limited age correlation, hindering population-level modeling. This study introduces a modified Poisson–Gauss (moPG) model, an enhancement of the Poisson–Gauss (PG) framework, to describe the human crystalline lens cross-sectional shape using solely age-related geometric parameters. The moPG extends the Poisson parameter m to real values via the Gamma function for continuous accommodation simulation and incorporates an equatorial curvature parameter k , inspired by Köller’s egg equation, for independent equatorial control. Analyzing pupil-limited conditions in OCT and MRI datasets, a non-zero k was suggested for biological plausibility. Evaluations on in vitro lens profiles ( n =10) showed better fitting than the PG model, with reduced dependence on the lens shape factor. moPG-derived surface areas, estimated from eigenlenses -model-derived geometric parameters, correlated strongly (Pearson correlation r =0.999, p 0 drives equatorial radius increase, while increasing m promotes progressive equatorial rounding.

Breast cancer represents the most frequently diagnosed cancer among women globally, with Vietnam experiencing a dramatic doubling of incidence over two decades while lacking comprehensive screening programs. This study proposes a fully automated deep learning pipeline integrating YOLOv8m detection, U-Net segmentation with ResNet152 backbone, and ResNet101 classification for breast ultrasound analysis. Training and validation utilize four heterogeneous public datasets (BUSI, BUS-BRA, BUS-UCLM, QAMEBI) through rigorous five-fold cross-validation. The detection module achieves 99.09% precision and 94.74% recall, while classification demonstrates 91.05% accuracy with ground-truth masks and 82.30% accuracy in fully automated mode. A Streamlit-based clinical application enables real-world deployment in resource-limited Vietnamese healthcare settings. The multi-dataset approach enhances generalizability across diverse populations and imaging conditions, positioning this framework as a practical AI-assisted tool for early breast cancer screening.

Conventional 2D thyroid ultrasound, while safe and accessible, is limited by operator dependency and the absence of volumetric data, compromising accurate morphological assessment. This study introduces a novel deep learning-based pipeline for automated thyroid segmentation and 3D reconstruction from ultrasound images, addressing these challenges. The pipeline was trained on a dataset of manually annotated ultrasound images from 10 healthy subjects acquired using the Samsung HS40 system. Leveraging MultiResUNet with multi-scale feature extraction, model robustness was strengthened through data augmentation and morphological post-processing. To ensure anatomical coherence in the reconstructed 3D models, automated isthmus alignment was subsequently applied. Evaluated using the Leave-One-Out Person strategy, MultiResUNet achieved a mean Dice coefficient of 93.6% and IoU of 87.8%, significantly outperforming UNet (Dice: 79.7%, IoU: 68.8%). The resulting segmentation masks enabled precise 3D thyroid reconstruction via slice stacking and isosurface extraction. This pipeline offers a reliable, automated solution for volumetric analysis, enhancing clinical decision-making and medical education. Despite its efficacy on a small dataset, future work will expand to pathological cases and diverse ultrasound systems to broaden clinical applicability. These findings highlight the potential of MultiResUNet-driven automation in advancing computer-assisted thyroid diagnostics.

This study proposes a low-cost, wearable multimodal system for comprehensive cardiovascular monitoring, integrating electrocardiogram (ECG), phonocardiogram (PCG), and photoplethysmogram (PPG) signals. By leveraging the complementary strengths of these modalities, the system extracts key physiological markers: pre-ejection period (PEP) from ECG–PCG fusion to assess myocardial electromechanical timing; pulse transit time (PTT) from PCG–PPG fusion to reflect arterial stiffness; and pulse arrival time (PAT) from ECG–PPG fusion to characterize heart–arterial coupling. Experimental results show that PEP is significantly associated with body mass index (BMI) but not age, suggesting that age-related changes primarily affect the arterial system rather than myocardial function. In contrast, both PTT and PAT demonstrate moderate negative correlations with age and BMI, supporting their relevance as noninvasive indicators of vascular aging. Additionally, PAT exhibits a significant sex-based difference, highlighting physiological disparities between male and female cardiovascular systems. Overall, the proposed fusion-based approach demonstrates technical feasibility and clinical potential for scalable, preventive cardiovascular healthcare, enabling early risk detection, continuous at-home monitoring, and improved long-term health management.

Sleep is essential for physiological and mental well-being. Sleep quality is determined by the pattern of sleep stages rather than the total sleep time. An ideal nighttime sleep pattern includes 4–5 complete sleep cycles, each sleep cycle typically starting with 3 stages of non-Rapid Eye Movement Stages (REM) sleep (N1, N2, N3) and ending with a stage of REM sleep, interspersed with wakefulness during sleep (Wake after sleep onset - WASO). A disrupted sleep cycle, such as being awakened during REM or deep sleep (N3), leads to discomfort upon awakening, known as sleep inertia. The optimal wake-up time is when awakening does not cause sleep inertia. This study aims to determine the optimal wake-up time based on the transition from other stages to N1 and N2 of NREM and WASO. The study will use available data sources such as the (Cyclic Alternative Patterns - CAP) and Sleep-EDF Database Expanded, as well as data collected from actual measurements on the Nicolet One system. After preprocessing, the collected data as extracted with 17 features. Feature selection was performed using an independent t-test, followed by classification using Naïve Bayes, K-Nearest Neighbor, Support Vector Machine, and Decision Tree models. The results showed that the Decision Tree model achieved the highest accuracy (83.42%), sensitivity (89.86%), and correct prediction on total wake-up group prediction (88.97%). The results achieved show that the optimal wake-up time with data has been labeled by sleep experts using the Decision Tree model, with the highest accuracy (83.42%), sensitivity (89.86%), and the F1 macro (88.97%).