Professor Gong Guosheng of the University of Hong Kong announces new coronary pneumonia AI results with 88% accuracy

Recently, Professor Nguyen Kwok-sung, Head of the Department of Statistics and Actuarial Sciences of the University of Hong Kong, jointly published a new paper with a number of other scholars introducing a new online diagnostic system for new coronary pneumonia (click on the original version of the paper). It is understood that the system for the diagnosis of new coronary pneumonia accuracy of 88%, AUC value of 93%, sensitivity of 86%, specificity 90%,

Professor Gong Guosheng of the University of Hong Kong announces new coronary pneumonia AI results with 88% accuracy

The participants in the study were Dr. Liu Bin, Assistant Professor of Statistics, Southwest University of Finance and Economics, Graduate Students Gao Xiaoxue, He Mengxuan, Liu Wei and Dr. Liu Bin’s colleague Lu Fengmao (Assistant Professor, School of Statistics, Southwest University of Finance and Economics).

Currently, the paper is under review, but the COVID-19 diagnostic system is online, free to use, and Python programs and data are fully open source (open source:

Professor Nguyen Said the University of Hong Kong has been at the forefront of world scientific research through the study of SARS outbreaks and various influenza viruses. During SARS in 2003, researchers in Shenzhen and Hong Kong jointly announced that THE PRECURSOR of SARS virus had been found in wild animals such as beavers.

Professor Gong Guosheng of the University of Hong Kong announces new coronary pneumonia AI results with 88% accuracy

Based on many years of research experience in biostatistics and clinical trials, from the end of January 2020, a team led by Professor Nguyen Kwok-sung began experimenting with some research on new coronary pneumonia, and CT-based image diagnosis is one of the tasks.

However, because there is no public CT image dataset, the team spends a lot of time looking for and marking open samples.

Later, medRxiv had a job of compiling a pre-printed paper on CT image analysis of new crown patients. The paper extracted 746 CT images of patients from preprints of medRxiv and bioRxiv articles and trained a neural network in the second classification of new crown patients.

However, the results show that the predictive effect has not yet reached the clinical standard.

One reason, according to Professor Gong Guosheng, is the small sample size, and another is that the CT image sample itself is not fully detailed labeling information. The biggest difference between this CT data and traditional medical imaging data is that each sample comes from a medical imaging paper.

In these articles, clinicians describe in detail the characteristics of chest CT lesions in new crown patients, and some also make a careful comparative analysis of the lesions characteristics of other common lung diseases.

Thus, in Professor Gong’s view, “this data, although the sample size is limited, is very informative and is a representative and high-value data set.” “

The researchers further focused on the text messages attached to the sample and found that 760 papers covered descriptions of five lesions (Lesions) of neo-coronary pneumonia, with one or more lesions appearing on each patient’s CT image. By analyzing the diagnosis of CT images of patients diagnosed with new crowns, these five lesions are the main criteria for the diagnosis of new coronary pneumonia in imaging.

Professor Gong Guosheng of the University of Hong Kong announces new coronary pneumonia AI results with 88% accuracy

So the team designed a Lesion-Attention deep neural network model (LA-DNN) based on THE CT image.

On the one hand, the model can distinguish the characteristics of new crown patients and non-new crown patients, on the other hand, the model’s “attention” is focused on the lesions area, that is, learning multi-label lesions, this is the team’s LA-DNN (Focus-Deep Neural Networks) model, as clinicians use CT images to judge the disease will focus on abnormal lesions and slightly over the normal area, the model at the same time training two tasks, coordination. Thus, the performance of the model has been significantly improved, and the indicators have reached the clinical standard.

At the same time, the team also used migration learning, that is, the use of pre-trained VGG, DenseNet and other neural networks as the backbone of the model network.

After the new corona CT image diagnostic system came online, the team continued to collect new samples, the online system training samples doubled the initial sample size, and regularly retrained the model, the online system results than the results in the paper.

Regarding the future direction of the results, Professor Gong said he hoped that medical personnel fighting the outbreak on the front line would use the system to share data and collaborate on research to help further test and improve the system.

“At present, the outbreak in China is under control, and many other countries and regions are still under great pressure to hope that the system can play a role in areas where the outbreak is still severe, reducing the burden of nucleic acid testing.” “