Research Project
Air Quality Monitoring for Vulnerable Groups in Residential Using Multiple Hazard Gas Detector
To overcome the limitations of using existing non-intelligent, slow-responding, deficient gas sensors, we proposed a novel artificial-intelligent-based multiple hazard gas detector (MHGD) system that is mounted on a motor vehicle-based robot that can be remotely controlled. MHGD can help vulnerable person group who have a relatively higher risk to suffer hazardous events than normal people to detect harmful gases more quickly and precisely. Our system is tested through experiments and the results shows that the designed MHGD system could achieve an acceptable accuracy.
Rapid Disease Detection System
In this study, we aimed to identify VOCs associated with lung cancer and perform targeted exhaled air analysis to detect lung cancer patients non-invasively. Most of the work was dedicated to the development of a novel lung cancer early detection system: the Rapid Disease Detection System (RDDS), which is different from the previous invasive system. RDDS can complete detection by analysing the patient's breath, which greatly simplifies the detection and analysis process. Therefore, the system is characterised by speed, non-invasiveness, ease of use and low detection costs.
Multimodal Learning for Non-small Cell Lung Cancer Prognosis
Fig 1: The architecture comparison of existing methods and our proposed Lite-ProSENet
This project focuses on building an adjuvant medical system for lung cancer prognosis to help clinicians to develop timely treatment plans and improve patients' quality of life. Here, the survival prediction model is at the heart of the system. Unlike previous methods that rely only on clinical text data for survival prediction, we assume that the information contained in visual data (e.g. CT images) and text data (e.g. textual clinical records) complement each other to form a more complete feature representation. On this basis, we develop the first multimodal network for NSCLC survival analysis, which takes deep learning-based NSCLC survival analysis one step forward by simultaneously considering the textual clinical data and the visual CT clues.
Experiments show that the proposed model can achieve better performance and provides a new state-of-the-art result of 89.3\% on the C-index.
Fig 2. A 3D model of NSCLC patient (from TCIA data archive).
Fig 3. A two-tower DNN model for learning similarity between textual clinical data and CT image representations.
Intelligent Person Re-identification Solutions for ShenZhen Airport
As one of China's most affluent cities, Shenzhen is home to nearly 18 million residents and is a vital shipping port as it connects China with neighbouring Hong Kong. While China continues to tackle new COVID -19 cases within its borders, Shenzhen is under pressure from the recent outbreak of the new Covid sub-variant. To contain transmission of COVID -19, authorities in Shenzhen have begun blanket searches of critical areas such as airports where COVID -19 positive patients and their close contacts are present. Currently, the searches are being conducted on a person-by-person basis. In the case of the airport, the staff of Shenzhen airport have to watch the video streams for hours or even days to find the targeted individuals. To alleviate this issues, we propose an intelligent person ReID system for Shenzhen Airport to reduce the high response time and alleviate the pressure of manual searching. Our human re-identification system has the advantages of high accuracy, fast processing speed and good usability. Our system can quickly access different types of video systems and perform real-time analysis and processing of full-target features as well as establish real-time indexing. The test results show that our system is able to track, position and accurately identify the target patient with their closely accompanying personnel. Now the system is being further improved to meet the needs of complex application scenarios such as communities, etc.