Research Publications
At AIVision Nexus Laboratory, we are dedicated to cutting-edge research in the field of deep learning. Our research focuses on solving practical imaging problems through the development of advanced deep learning models. Additionally, we create AI-powered modeling and forecasting systems for practical applications, and integrate Explainable AI in real-world decision-making systems. Below you will find a list of some of our latest research publications.
Z. Huang, R. Zhao, F.H.F. Leung, S. Banerjee, K.M. Lam, Y.P. Zheng, *S.H. Ling, “Landmark localization from medical images with generative distribution prior,” IEEE Transactions on Medical Imaging, 2024.
In medical image analysis, anatomical landmarks usually contain strong prior knowledge of their structural information. In this paper, we propose to promote medical landmark localization by modeling the underlying landmark distribution via normalizing flows. Specifically, we introduce the flow-based landmark distribution prior as a learnable objective function into a regression-based landmark localization framework. Moreover, we employ an integral operation to make the mapping from heatmaps to coordinates differentiable to further enhance heatmap-based localization with the learned distribution prior. Our proposed Normalizing Flow-based Distribution Prior (NFDP) employs a straightforward backbone and non-problem-tailored architecture (i.e., ResNet18), which delivers high-fidelity outputs across three X-ray-based landmark localization datasets. Remarkably, the proposed NFDP can do the job with minimal additional computational burden as the normalizing flows module is detached from the framework on inferencing. As compared to existing techniques, our proposed NFDP provides a superior balance between prediction accuracy and inference speed, making it a highly efficient and effective approach. The source code of this paper is available at https://github.com/jacksonhzx95/NFDP.
S. Banerjee, Z. Huang, J. Lyu, F.H.F. Leung, T.T.Y. Lee, D. Yang, Y.P. Zheng, J. McAviney, and *S.H. Ling, “Automatic assessment of ultrasound curvature angle for scoliosis detection using 3D ultrasound volume project imaging,” Ultrasound in Medicine & Biology, 2024.
Objective
Scoliosis is a spinal deformation in which the spine takes a lateral curvature, generating an angle in the coronal plane. The conventional method for detecting scoliosis is measurement of the Cobb angle in spine images obtained by anterior X-ray scanning. Ultrasound imaging of the spine is found to be less ionising than traditional radiographic modalities. For posterior ultrasound scanning, alternate indices of the spinous process angle (SPA) and ultrasound curve angle (UCA) were developed and have proven comparable to those of the traditional Cobb angle. In SPA, the measurements are made using the spinous processes as an anatomical reference, leading to an underestimation of the traditionally used Cobb angles. Alternatively, in UCA, more lateral features of the spine are employed for measurement of the main thoracic and thoracolumbar angles; however, clear identification of bony features is required. The current practice of UCA angle measurement is manual. This research attempts to automate the process so that the errors related to human intervention can be avoided and the scalability of ultrasound scoliosis diagnosis can be improved. The key objective is to develop an automatic scoliosis diagnosis system using 3-D ultrasound imaging.
Methods
The novel diagnosis system is a three-step process: (i) finding the ultrasound spine image with the most visible lateral features using the convolutional RankNet algorithm; (ii) segmenting the bony features from the noisy ultrasound images using joint spine segmentation and noise removal; and (iii) calculating the UCA automatically using a newly developed centroid pairing and inscribed rectangle slope method.
Results
The proposed method was evaluated on 109 patients with scoliosis of different severity. The results obtained had a good correlation with manually measured UCAs (R2=0.9784 for the main thoracic angle and R2=0.9671 for the thoracolumbar angle) and a clinically acceptable mean absolute difference of the main thoracic angle (2.82 ± 2.67°) and thoracolumbar angle (3.34 ± 2.83°).
Conclusion
The proposed method establishes a very promising approach for enabling the applications of economic 3-D ultrasound volume projection imaging for mass screening of scoliosis.
Z. Namadchian, A. Shoeibi, A. Zare, J.M. Gorriz, H.K. Lam, and S.H. Ling, “Stability Analysis of Dynamic General Type-2 Control System with Uncertainty,” IEEE Trans. On Systems, Man and Cybernetics: Systems, vol. 54, no. 3, Mar. 2024.
A growing body of literature has proved that general type-2 fuzzy control systems (GT2 FCSs) are reliable, robust, and safe control systems against severe external disturbances, high levels of noise, and uncertainty that are inevitable in real-world applications. Further studies on the GT2 FCSs are therefore needed to provide new insights into these control systems, in particular over their stability and computational complexity problems. Unlike the existing works on stability analysis of GT2 FCSs in the time domain, the aim of this article is to assess the stability properties of these systems in the frequency domain through a novel intuitive method. Before delving into stability analysis, an initial step involves reducing computational complexity. This is achieved by introducing a streamlined version of the GT2 fuzzy controller (GT2 FC). This simplified architecture is constructed using a series of zSlices-based interval type-2 fuzzy-logic controls (zIT2 FLCs) situated at specific zLevels. Each zIT2 FLC is composed of two embedded type-1 fuzzy FLCs (T1 FLCs). The subsequent phase entails a methodical procedure for assessing stability based on the existence of limit cycles. The initial stage involves the linearization of the simplified GT2 FC through the derivation of its describing function (DF). Next, a combination of the parameter plane approach and the particle swarm optimization (PSO) technique is leveraged. These techniques serve the purpose of pinpointing the regions corresponding to limit cycles and asymptotic stability with a focus on improving the stability boundary. Following this, the analysis shifts toward quantifying the system’s resilience in the face of uncertainty. Stability margins required to generate a limit cycle are calculated using stability equations. This step provides a measure of how robust the closed-loop system remains under varying degrees of uncertainty. Finally, three simulation examples are presented to justify the advantages of the proposed approach.
W. Tun, J. Wong, and S.H. Ling*, “Advancing Fault Detection in HVAC Systems: Unifying Gramian Angular Field and 2D Deep Convolutional Neural Networks for Enhanced Performance,” Sensors, vol. 23, p.7690, 2023.
Efficiency and comfort in buildings rely on on well-functioning HVAC systems. However, system faults can compromise performance. Modern data-driven fault detection methods, considering diverse techniques, encounter challenges in understanding intricate interactions and adapting to dynamic conditions present in HVAC systems during occupancy periods. Implementing fault detection during active operation, which aligns with real-world scenarios and captures dynamic interactions and environmental changes, is considered highly valuable. To address this, utilizing the dynamic simulation system HVAC SIMulation PLUS (HVACSIM+), an HVAC fault model was developed using 194 sensor signals from each HVAC component within a single-story, four-room building. The advanced HVAC fault detection framework, leveraging simulated HVAC operational scenarios with the Gramian angular field (GAF) and two-dimensional convolutional neural networks (GAF-2DCNNs), offers a robust and proactive solution. By utilizing the GAF capacity to convert time-series sensor data into informative 2D images, integrated with 2DCNN for automated feature extraction, hidden temporal relationships within 1D signals are captured. After training on nine significant HVAC faults and normal conditions during occupancy, the effectiveness of the proposed GAF-2DCNN is evaluated through comparisons with support vector machine (SVM), random forest (RF), and hybrid RF-SVM, one-dimensional convolutional neural networks (1D-CNNs). The results demonstrates an impressive overall accuracy of 97%, accompanied by precision, recall, and F1 scores that surpass 90% for individual HVAC faults. Through the introduction of the unified approach that integrates HVACSIM+ simulated data and GAF-2DCNN, a notable enhancement in robustness and reliability for handling substantial HVAC faults is achieved.
Y. An, J.K.W. Wong, and S.H. Ling, “Development of Real-time Brain-Computer Interface Control System for Robot,” Applied Soft Computing, 2024.
Electroencephalogram (EEG)-based brain-computer interfaces (BCI) have been considered a prevailing non-invasive method for collecting human biomedical signals by attaching electrodes to the scalp. However, it is difficult to detect and use these signals to control an online BCI robot in a real environment owing to environmental noise. In this study, a novel state recognition model is proposed to determine and improve EEG signal states. First, a Long Short-Term Memory Convolutional Neural Network (LSTM-CNN) was designed to extract EEG features along the time sequence. During this process, errors caused by the randomness of the mind or external environmental factors may be generated. Thus, an actor-critic based decision-making model was proposed to correct these errors. The model consists of two networks that can be used to predict the final signal state based on both the current signal state probability and past signal state probabilities. Subsequently, a hybrid BCI real-time control system application is proposed to control a BCI robot. The Unicorn Hybrid Black EEG device was used to acquire brain signals. A data transmission system was constructed using OpenViBE to transfer data. An EEG classification system was built to classify the BCI commands. In this experiment, EEG data from five subjects were collected to train and test the performance and reliability of the proposed control system. The system records the time spent by the robot and the moving distance. Experimental results were provided to demonstrate the feasibility of the real-time control system. Compared to similar BCI studies, the proposed hybrid BCI real-time control system can accurately classify seven BCI commands in a more reliable and precise manner. Overall, the offline testing accuracy was 87.20%. When we apply the proposed system to control a BCI robot in a real environment, the average online control accuracy is 93.12%, and the mean information transmission rate is 67.07 bits/min, which is better than those of some state-of-the-art control systems. This shows that the proposed hybrid BCI real-time control system demonstrated higher reliability, which can be used in practical BCI control applications.