BEIJING – More and more companies and researchers in China are rolling out artificial intelligence (AI)-based systems that can process hundreds of computed tomography (CT) images in seconds to speed up diagnosis of COVID-19 and assist in its containment.
China’s national guidelines have recommended CT scans as a key method in diagnosing COVID-19. While pathogenic laboratory testing is regarded as the diagnostic gold standard, it takes a much longer time and can generate false positive results. The epidemic has driven companies and researchers to come up with more efficient ways to cope with the huge demand for diagnosis.
Ping An Smart Healthcare, a team under Ping An Insurance (Group) Company of China Ltd.’s subsidiary Ping An Smart City, is one of the latest to announce their process. On Feb 28, it unveiled the COVID-19 smart image-reading system that it claims can generate analysis results in about 15 seconds with an accuracy rate above 90%, whereas radiologists can spend up to 15 minutes reading the CT images of a patient suspected of contracting COVID-19.
The health care team of the Chinese insurer said the AI analysis engine can conduct a comparative analysis of multiple CT scan images of the same patient while measuring the changes in lesions. The system also helps doctors track the development of the disease, evaluate treatment and make prognoses for patients.
“Patients with COVID-19 need multiple CT scans during the treatment. Comparing multiple images is a time-consuming task and it cannot be accurately completed manually,” said Ping An’s chief scientist Xiao Jing.
He added that the system can effectively improve diagnostic accuracy and doctors’ image-reading efficiency. Ultimately, Ping An aims to help doctors diagnose, triage and evaluate COVID-19 patients faster and more effectively.
Geoff Kau, co-president and chief strategy officer of Ping An Smart City, said more than 1,500 medical institutions have used the system on more than 5,000 patients since Feb 19.
Several companies are also using AI for the same task.
AI-focused Beijing Infervision Technology Co. Ltd. developed Inferread CT Pneumonia software to detect lesions from possible pneumonia caused by the coronavirus with analyses of its volume, shape and density, while comparing changes in multiple lung lesions from the CT image. The whole process can take as little as 10 seconds.
The AI startup’s software is said to have been used at 34 hospitals in China and reviewed over 32,000 cases. Matt Deng, chief scientist and director of Infervision North America, said the software is highly accurate with 95% sensitivity.
The software was used to detect cancer in lung CTs before being repurposed for COVID-19.
“The quick acknowledgement of the quantitative nature, along with AI’s effective patient triage, enabled doctors to make fast and confident decisions on the pneumonia-probable cases, especially when they need to read thousands of cases a day,” said Infervision in a statement.
“With every second saved, the COVID-19-probable patients can be directed into further procedures quicker without spending hours waiting at the hospital and posing a severe risk of cross-infection with others,” it added.
Ping An and Infervision are not alone in this cause.
Covering the same area, Alibaba’s tech unit DAMO Academy launched an AI system to analyze CT images within 20 seconds with an accuracy rate of 96%, AI startup Deepwise Technology Co. Ltd. rolled out the Dr. Wise cloud+whole lung AI-aided medical diagnosis system, AI firm Iflytek Science and Technology Co. Ltd. jointly developed an AI-based COVID-19 diagnosis platform with the Chinese Academy of Sciences, and Shanghai Yitu Information Technology Co. Ltd. introduced a scan-reading system.
In academia, efforts are being made by Chinese researchers, who have published in medRxiv to demonstrate how deep learning can enable AI-assisted diagnosis to achieve results much faster.
In one study, researchers collected hundreds of CT images of pathogen-confirmed COVID-19 cases along with those previously diagnosed with typical viral pneumonia, then modified the inception migration-learning model to establish the algorithm, followed by internal and external validation.
They said the internal validation achieved a total accuracy of 82.9% with specificity of 80.5% and sensitivity of 84%, and the external testing dataset showed a total accuracy of 73.1% with specificity of 67% and sensitivity of 74%. They say the results are the proof-of-principle for using artificial intelligence to extract radiological features for timely and accurate COVID-19 diagnosis.
In another study, researchers reached a similar conclusion after going through over 46,000 images. They said the deep learning model showed comparable levels of performance with expert radiologists, and greatly improve the efficiency of radiologists in clinical practice.
AI experts are not at all surprised to see a huge interest in leveraging AI to fight COVID-19.
“There's a lot of existing work on automated interpretation of chest imaging that knows how to deal with patient variation and movement between imaging sets. …We have known techniques to deal with that,” David Rawlinson, a Melbourne-based AI expert who has co-authored a number of AI-related papers, told BioWorld.
He added that the requirements for detecting opacity in lung imaging are already met, so dealing with COVID-19 doesn't require many changes.
“The upper bound on the utility of any automation is whether the gold standard – human expert reporting – is actually conclusive or not. It seems like the RT-PCR and imaging are both imperfect, and some combination of both is the most reliable method,” he added.