AHAudrey Huong

Audrey Huong

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article

Previous Organizations / Institutes

The organizations / institutes from the scholar's other articles in the past ten years


Audrey Huong has a total of 86 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    A Website-Based Classification of Cervical Spine Conditions Using Convolutional Neural Network.

    by Ee Herng Loh, Kim Gaik Tay, Mohd Norzali Haji Mohd, Audrey Huong

    Accurate assessment of cervical spine alignment is essential for diagnosing and monitoring cervical spine disorders. However, manual interpretation of cervical spine X-ray images is time-consuming and prone to variability. To address this, the study proposes an automated web-based prototype that uses a convolutional neural network (CNN) and a rule-based Cobb angle classification for cervical spinal conditions (lordosis, kyphosis, and straight neck). A stratified fivefold cross-validation (CV) framework with separate train, validation, and test splits per fold, incorporating transfer learning (TL), fine-tuning (FT), and hyperparameter optimisation (HPO) using Optuna, was employed. Two pre-trained CNNs, ResNet152V2 and ConvNeXt Tiny, were evaluated using the same data splits. Results showed that ResNet152V2 outperformed ConvNeXt Tiny, achieving a mean accuracy of 86.0 ± 1.21% (95% bootstrap confidence intervals [CI] 84.2%-87.9%), demonstrating a greater stability with a lower standard deviation (SD) and a narrower CI. Both the image-level paired permutation test and the McNemar test showed a statistically significant performance difference of 4.47%. Despite ResNet152V2's heavier architecture, higher GPU latency, and lower GPU throughput, the practical processing time difference is minor because the computation runs on a backend server rather than on a resource-constrained mobile platform. Therefore, the optimised ResNet152V2 model was selected for deployment via a Gradio web interface hosted on Hugging Face. In addition, the CNN-based classification significantly outperformed the Cobb angle-based classification (96.67% vs 88.33% accuracy). A 95% bootstrap CI with 10,000 resamples showed a significant improvement in CNN-based classification over the rule-based classification. A right-tailed test yielded a p-value of 0.0292 at the 0.05 significance level, consistent with 97.08% of bootstrap samples showing a positive difference in accuracy. Ultimately, this study demonstrates the feasibility of integrating an optimised deep learning model and Cobb-angle classification into a real-time cervical spine research prototype. Future work will focus on larger datasets and lightweight model designs to enhance its applicability.

  • An Automatic PSO-Optimized Spatial-Temporal Neural Network for Blood Perfusion Analysis

    by Audrey Huong, Hui Ling Chua, Ser Lee Loh, Kim Gaik Tay, Kok Beng Gan, Xavier Ngu

    Existing technologies for evaluating oxygen delivery efficiency and blood perfusion are limited to complex equipment, whilst accurate diagnosis of compromised microvascular circulation relies on domain-specific mathematical models. Although an ultrasonic flaw detector can be an economical alternative, developing an optimal prediction model for a specific imaging task requires manual fine-tuning and customization. This study aims to introduce an optimization-empowered hybrid spatial-temporal network design framework for rapid and nondestructive blood perfusion status classification using a simple photoacoustic system. To examine this framework, this paper conducted experiments on seventeen volunteers. External intervention was performed to modify the perfusion of the chosen body part; the oxygenation-related photoacoustic (PA) signal was recorded after the chosen site was illuminated with temporally modulated light beams of wavelengths 500 nm and 550 nm. Meanwhile, the network design and training processes used the Particle Swarm Optimization (PSO) method for efficient model convergence on the small PA dataset. Experimental results comparing the designed hybrid model against two benchmark networks trained for the same microcirculatory perfusion disturbance problem confirmed the competitiveness of the proposed model with a superior mean classification accuracy, specificity, precision, and F 1 score of 80.8 %, 92.3 %, 89.9 %, and 79.4 %, respectively. This is further supported by the result of the Receiver operating characteristic (ROC) curve, showing a good area under the curve (AUC) of 0.91. These findings offered evidence of hemodynamic fluctuations at the microcirculatory level to compensate for low oxygen under ischemic-hypoxic conditions. This study concluded that the future of this system includes its industrial and medical applications for nondestructive inspections, home use, or bedside microcirculation monitoring to optimize wound and postoperative outcomes.

  • Open Access

    Cascaded Segmentation-Classification Framework Optimized for Sports Action Recognition Using Still Images: A Comparative Analysis of PSO and Grid Search

    by Audrey Huong, Ser Lee Loh, Kok Beng Gan, Xavier Ngu

    The conventional method of evaluating human actions and activities requires integrating complex hardware and interpretation systems. This article proposes using a hybrid segmentation-classification system to recognize human sports action automatically based on still frames. This study tested the proposed framework on a public dataset containing images of humans performing three athletic actions: dancing, performing martial arts, and playing net sports. This framework combined a two-stage U-Net and EfficientNet-B0, and GoogleNet, and is optimized using particle swarm optimization (PSO) and grid search for cascaded segmentation and classification problems. Results indicated that PSO improves segmentation and classification accuracies by 5 % and 60 %, respectively, compared to the conventional grid search. The PSO segmentation results showed good agreement between the predicted mask and its ground truth, with overlapping scores of 0.85-0.9. The consecutive classification experiments revealed a slight superiority in the performance of EfficientNet-B0 with an accuracy of 0.86-0.88 over GoogleNet (~0.82-0.84), which also showed a lower convergence efficiency. While no significant difference was observed in the recall scores of EfficientNet between the weighted and unweighted loss methods, GoogleNet showed a considerable improvement in the true positive rates from 0.6177 to 0.7160 using the weighted loss strategy. Despite using the weighted loss method, there were negligible improvements in the performance of grid search-optimized networks. The poor classification results suggest low model generalization ability using manual tuning. This study concluded that the PSO-optimized cascaded segmentation-classification framework could potentially leverage advancements in human movement evaluation and rehabilitation assessment for sports applications. This system can be improved by adopting larger datasets or collaborative machine learning training to enhance model convergence for practical rehabilitation applications to assess users’ physical and fitness performance.

  • Open Access

    Water Level and Flow Detection System: An IoT-Based Flood Monitoring Application

    by Syed Zainul Shafiq Syed Zaifudin, Wan Mahani Hafizah Wan Mahmud, Audrey Huong, Nur Anida Jumadi, Raja Mohd Aizat Raja Izaham, Hong-Seng Gan

    In Malaysia, floods have become increasingly prevalent, causing significant damage to various aspects of life including infrastructure, property and the economy. The risk to human lives is particularly high in low-lying areas and areas with poor drainage systems. Water contamination during floods can also lead to health issues and the spread of waterborne diseases. The psychological impact of floods should not be underestimated, as individuals may experience a range of emotions such as frustration, sadness, fear and anger. To address these challenges, this study focuses on developing a flood alert system using Internet of Things (IoT) technology. By utilizing ESP32 microcontroller, flow sensor, float sensors, water flow as well as water level can be accurately measured and users can be alerted through the Blynk application. The system ensures accuracy in measuring flow rates, with recorded values ranging from 2.404 L/min to 2.444 L/min. This device can measure water flow and water level which will be generated through ESP32 microcontroller and notifying the user via the Blynk application as for the IoT part. This project has been carried out in indoor experiments to evaluate the efficiency of the float and flow sensors, respectively. Through indoor experiments, it has been demonstrated that this system is capable of effectively monitoring and alerting users to potential flood incidents, helping to safeguard lives and property.


  • Open Access

    Optical Simulation Study for Non-Invasive Jaundice Application: Effect of Bilirubin, Melanosome, LED Radial Width Distribution and Optode Distance

    by Nur Safwah Abd Rahaman, Nur Anida Jumadi, Audrey Huong Kah Ching, Wan Mahani Hafizah Wan Mahmud, Junaid Ali

    One of the promising diagnosis devices is the non-invasive bilirubin detection for estimating newborn jaundice. From the light interaction through tissue perspective, less attention has been given to the effect of LED radial width distribution at different melanosome content and bilirubin concentrations. Besides, different optode (emitter-detector) spacing can affect the light measured at the detector. This paper aims to investigate the effects of three LED radial width distributions (30°, 40° and 60°) on three types of skin pigmentation (light-skinned Caucasians, well-tanned Caucasians and Mediterranean, and dark-skinned pigmented Africans) based on hyperbilirubinemia and excessive hyperbilirubinemia conditions by using MCmatlab to simulate a 3D cuboid shape of neonatal forehead tissue layers. To mimic the tissues model, optical properties such as absorption coefficient (μ_a), reduced scattering coefficient (〖'μ〗_s), and scattering anisotropy (g) were applied in every layer. Wavelength 470 nm was chosen since it is the most responsive wavelength at bilirubin’s absorption spectra. Two optode spacing (0.7 cm and 2 cm) were used to demonstrate the effect of LED radial width. Main findings revealed that for all three types of skin, 30° LED radial width distribution resulted different trends in 0.7 cm and 2 cm of optode spacing. At 0.7 cm, a higher percentage of incident light was collected at the detector, whereas diffuse reflectance was lower. However, at a 2cm distance, both diffuse reflectance and the percentage of incident light collected on the detector were lower. These measured outcomes in diffuse reflectance on light skin were increased by six times, and the incident light collected was one time larger than on tanned and dark skin, as light skin colour absorbs less light and therefore gives a better reflection. Different optode spacing will be investigated for future studies as this information can serve as valuable input when designing optical hardware for non-invasive jaundice detection.

  • Open Access

    Optical Coherence Tomography Image Analysis for Detection of Alzheimer’s Disease: A Comprehensive Structured Review

    by Wan Mahani Hafizah Wan Mahmud, Audrey Kah Ching Huong, Nur Anida Jumadi, Raja Mohd Aizat Raja Izaham, Hong-Seng Gan

    Optical Coherence Tomography (OCT) has emerged as a promising non-invasive imaging modality for the early detection of Alzheimer’s Disease (AD). This systematic literature review aims to consolidate current research on OCT image analysis for AD detection, addressing the growing need for early and accurate diagnostic tools. Despite the advances in neuroimaging, early diagnosis of AD remains challenging due to its asymptomatic nature in initial stages and the invasiveness of traditional methods. To achieve this, we conducted an extensive search of related articles from reputable databases (Scopus and Web of Science), focusing on studies published between 2022-2024. The flow of study was based on PRISMA framework. The database found (n = 29) final primary data. This review was divided into three themes, (1) retinal and ocular biomarkers for AD, (2) optical coherence tomography angiography (OCTA) and imaging techniques, and (3) machine learning and computational approaches for Alzheimer’s disease diagnosis. Key findings include the enlargement of the periarteriole capillary-free zone and changes in retinal nerve fibre layer thickness as potential AD biomarkers. Based on the review, the implementation of image analysis for OCT images have shown substantial potential for AD detection. By evaluating the past studies, gaps for the current research were discovered including the need for larger, more diverse cohorts and longitudinal studies to validate these biomarkers. In summary, detection of AD is possible through thorough OCT image analysis, but further research could be suggested to enhance its clinical applicability and reliability.


  • Open Access

    Android-Based App Guava Leaf Diseases Identification using Convolution Neural Network

    by Choon Fai Foo, Kim Gaik Tay, Osamah Al-Qershi, Audrey Huong, Chang Choon Chew, Shehab Abdulhabib Alzaeemi

    Guava is an economically significant fruit crop in many regions. Conventional plant disease identification depends on skilled workers visually inspecting samples, which can be laborious and error-prone. Therefore, the goal of this study is to create an Android app that uses InceptionV3 to precisely determine the health of guava plants, with an emphasis on hyperparameter tuning and ideal model selection. Inception V3 was chosen as the base model for this study as it outperformed the 25 pre-trained models from Keras API. The Guava leaves dataset consists of 120 raw images collected from real-time capture and the Google website. The images belong to four categories: Algal Spot, Red Rust, Whitefly, and Healthy leaf. The dataset was split into 60% for training and 40% for validation. To increase the number of images, augmentation was performed to multiply each training by 5 times. The 2048 features were extracted from the second last layer of Inception V3. These features were saved in the Numpy format to facilitate hyperparameter tuning. The best hyperparameter tuning results were achieved with a batch size of 4, a learning rate of 0.001, and the Adamax optimizer for 50 epochs, resulting in a validation accuracy of 0.99. This set of hyperparameters was used to train the model, yielding in a training accuracy of 0.98 and a validation accuracy of 0.9. The trained model was exported in the tflite file format and integrated into an Android app developed using Android Studio. To evaluate the performance evaluation of the app, 15 unseen images per class were used and the app achieved an overall F1-Score of 0.85, a recall of 0.85, a precision of 0.89, and a test accuracy of 0.85.


  • Open Access

    Blood CO Status Classification Using UV-VIS Spectroscopy and PSO-optimized 1D-CNN Model

    by Audrey Huong, Kim Gaik Tay, Kok Beng Gan, Xavier Ngu

    Rapid and effective blood carbon monoxide (CO) assessment is of great importance, especially in estimating CO-related morbidity and instituting effective preventive measures. The conventional detection methods using CO breath analysis lack sensitivity, while collecting biological fluid samples for CO level measurement is prone to external contamination and expensive for frequent use. This study proposes a one-dimensional convolutional neural network (1D-CNN) consisting of three stacked biconvolutional layers for binary classification of blood CO status using the diffuse reflectance spectroscopy technique. Iterative particle swarm optimization (PSO) has efficiently found the best network parameters to learn important features from the reflectance spectroscopy data. The findings showed good testing accuracy, specificity, and precision of 92.9%, 90%, and 89.7%, respectively, and a high sensitivity of 96.3% in determining abnormal blood CO among smokers using the proposed CNN network. Comparisons with eight existing machine learning and deep learning models revealed the proposed method’s effectiveness in classifying blood CO status while reducing computing time by 8–13 folds. The findings of this work provide new insights that are valuable for researchers in neural network design automation, healthcare management, and skin-related research, specifically for application in nondestructive evaluation and clinical decision-making.

  • MI-OPTNET: AN OPTIMIZED DEEP LEARNING FRAMEWORK FOR MYOCARDIAL INFARCTION DETECTION

    by Audrey Huong, Kim Gaik Tay, Kok Beng Gan, Xavier Ngu

    The conventional means of myocardial infarction (MI) detection using a 12-lead electrocardiogram (ECG) system include a pretrained network and machine learning interpretation of the complex ECG signals. They are computationally inefficient and demand high-performance hardware. Here, for the first time, we introduce an effective framework (MI-OptNet) using the particle swarm optimization model (PSO) in the design of a lightweight hybrid network combining convolutional neural network (CNN)-long short terms memory (LSTM) for MI and normal ECG detection. We optimized important design and training parameters based on limb leads’ signals and identified leads III and VI as the best ECG leads for the task based on their high classification performance ranging between 80 – 90 %, suggesting that they may provide more information about MI than the others. The other strategy of fusing the scores from all models at the decision level achieved the best result with a 10 % increase in the evaluated metrics. Our findings support the flexibility and adaptability of our framework for the design process using minimal computer efforts. We concluded that this approach may be used for other classification problems to assist engineers and designers in efficient decision-making and to solve complex signal classification and recognition problems.


  • Open Access

    Dorsal Hand Vein Pattern Recognition: A Comparison between Manual and Automatic Segmentation Methods

    by Waheed Ali Laghari, Audrey Huong, Kim Gaik Tay, Chang Choon Chew

    Objectives: Various techniques for dorsal hand vein (DHV) pattern extraction have been introduced using small datasets with poor and inconsistent segmentation. This work compared manual segmentation with our proposed hybrid automatic segmentation method (HHM) for this classification problem.Methods: Manual segmentation involved selecting a region-of-interest (ROI) in images from the Bosphorus dataset to generate ground truth data. The HHM combined histogram equalization and morphological and thresholding-based algorithms to localize veins from hand images. The data were divided into training, validation, and testing sets with an 8:1:1 ratio before training AlexNet. We considered three image augmentation strategies to enlarge our training sets. The best training hyperparameters were found using the manually segmented dataset.Results: We obtained a good test accuracy (91.5%) using the model trained with manually segmented images. The HHM method showed slightly inferior performance (76.5%). Considerable improvement was observed in the test accuracy of the model trained with the inclusion of automatically segmented and augmented images (84%), with low false acceptance and false rejection rates (0.00035% and 0.095%, respectively). A comparison with past studies further demonstrated the competitiveness of our technique.Conclusions: Our technique can be feasible for extracting the ROI in DHV images. This strategy provides higher consistency and greater efficiency than the manual approach.