Researchers at The Chinese University of Hong Kong (CUHK) have developed an artificial intelligence (AI)-based system for the automated, rapid and accurate detection of COVID-19 infections in chest computed tomography (CT) images.
The system can provide immediate results, without the need for clinicians to interpret images. It could be potentially used in radiology imaging departments in hospitals.
The CUHK team already have put the system to the test. Experimental outcomes have revealed that the AI model managed to yield a competitive performance in lesion detection in comparison with radiologist interpretation of chest CT scans across local, regional and global patients.
Its wide validation and applicability on cohorts with various imaging scanners and different demographics show that the established AI model holds great potential in complex real-world situations.
“The established AI system is validated on multiple, unseen, independent external cohorts from mainland China and Europe, showing the potential and feasibility to build large-scale medical datasets with privacy protection, so as to rapidly develop reliable AI models amidst a global disease outbreak such as the COVID-19 pandemic,” said Dou Qi, an assistant professor at CUHK’s Department of Computer Science and Engineering, who co-led the multidisciplinary research team.
“We aggregate data from multiple hospitals in Hong Kong, so that the model is robust to various data distributions, and applicable to other unseen hospitals in different regions and countries,” Qi told BioWorld.
Heng Pheng Ann from CUHK’s department of computer science and engineering, believes this can lift the burden from clinicians in interpreting images across multiple data pools.
“Making use of the cutting-edge federated learning techniques, the AI system can effectively coordinate the patient data across multiple clinical centers in Hong Kong, including Prince of Wales Hospital, for model development,” said Ann, who also led the team.
“Given the unavoidable challenge of data heterogeneity in medical images, a multicenter collaborative effort is essential to capture diverse data distributions, which enhances model reliability for unleashing the potential of AI-powered medical image diagnosis in complex clinical practice.”
The AI system is trained on multicenter data in Hong Kong using what the researchers call “new federated learning techniques” without the need to centralize data in one place, thus protecting patient privacy.
Given how fast pandemics like COVID-19 progress, there is often not enough time to set up complicated data sharing agreements across institutions or even countries. Thus, privacy is an important factor in such data driven diagnostics.
“Privacy preserving machine learning acts as an important enabler under such situations to gather efforts on digital medicine technology for providing reliable clinical assistance for timely patient care,” Simon Yu, a professor and chairman of CUHK’s department of imaging and interventional radiology at the faculty of medicine.
He believes the study for the AI system demonstrates the feasibility and efficacy of federated learning for COVID-19 image analysis, where collaborative effort is especially valuable at a time of global crisis.
“More importantly, beyond assisting COVID-19 management, we believe that AI, which protects patient privacy and achieves reliable generalizability in practice, has enormous potential to revolutionize smart hospitals and health care systems in Hong Kong and worldwide,” said Yu, a co-leader in its development.
At the moment, radiological imaging plays a complementary role with clinical diagnostic testing in COVID-19 diagnosis in most countries. It can, and is, used to effectively assess the severity and progression of an infection.
With the large amount of data analysis and medical image interpretations needed, automated diagnostic methods with AI would be especially useful in facilitating effective management of the pandemic.
“Besides a high diagnostic accuracy, the AI system can also present a remarkable speed advantage to clinician interpretation,” said Tiffany So, assistant professor from CUHK’s department of imaging and interventional radiology at the faculty of medicine.
“In traditional clinical diagnosis, review and interpretation of a single chest CT takes at least 5 to 10 minutes for clinicians. In contrast, the AI system can accurately evaluate the same CT data in around 40 minutes, showing immense potential to support real-time clinical practice,” added So, a team leader on this discovery.
For now, wider validation on the model is planned and there is no specific timeline for bringing this to market.
The research team believes the AI system can also be expanded to general lung disease lesion detection, such as lung cancer and other lung infections like pneumonia.