Artificial neural networks help tell COVID-19 from other types of pneumonia.
Reading chest scans is a toilsome task, considering the fact that each patient has 400 scans and medics are also needed at sickbeds to take care of patients. This is where AI steps in.
On Feb. 21, a diagnostic system assisted by artificial intelligence (AI) was installed at Shanghai Public Health Clinical Center (SPHCC) affiliated to Fudan University, where all confirmed COVID-19 cases in Shanghai are treated. The AI system is developed by a team of researchers from Fudan School of Computer Science, Fudan Institute of Big Data, and the Department of Radiology at SPHCC.
A researcher installs the AI system at SPHCC
Fudan’s AI diagnostic system is able to identify and classify types of diseases based on chest CT scans, telling COVID-19 from other viral or bacterial pneumonia. Compared to nucleic acid test, a widely used approach to diagnose infection, which returns a false-negative rate at 30%-50%, Fudan’s AI system reports such an occurrence as low as 7%.
A: CT chest scan; B: Visual feedback from the AI system
The advantage of AI over the human eye, of course, lies in its efficiency. With Fudan’s AI system, lesions can be located within seconds while it takes a human radiologist 5-10 minutes.
But it is the superior algorithms inside that makes this system stands out from its counterparts developed by tech firms and research institutes across China.
According to Prof. Xue Xiangyang from Fudan School of Computer Science, the AI system is designed with two algorithms, one for classification of different types of pneumonia and the other detection of lung lesions.
To build any AI model, the first step is to collect the data needed to train the neural network. However, during the early period of the outbreak, it was impossible to gain a huge set of scans from patients. Also, labeling lesions takes a lot of time out of medics when they are busy treating patients. The team thus resorted to Small Sample Learning (SSL), which is suitable for network training with a small amount of data, said Xue.
Currently, a major problem with COVID-19 diagnosis is accuracy, as other types of pneumonia may lead to similar visuals, such as ground-glass opacities. However, Xue said, the location of lesions in the lung has been confirmed as a vital indicator to differentiate COVID-19 from other types of pneumonia.
The various sizes of lesions pose another challenge to the development of the system. To solve this problem, the team introduced a hierarchical neural network. Each layer of the network corresponds with a certain size of the infected area. “The underlying layer concentrates on details, which means it is sensitive to small lesions, while the middle and top layers, concentrate on bigger lesions. By putting different weights on each layers, eventually we found the solution to detect all lesions at the same time regardless of their sizes,” said Xue.
He also stressed that in the case of a tiny lesion or a complicated scan result which the network had never encountered, human doctors would still be the ones to make the call based on their clinical experience.
“During test runs, we used a case where the patient experienced a minor relapse after treatment. The pneumonia classification algorithm diagnosed the patient as healthy, while the lesion detection algorithm reported the relapse,” Xue recalled. To improve accuracy, the team designed an integrated-analysis function that could make a comprehensive assessment based on the outcomes of multiple algorithms.
Team members work at home
Xue said he looked forward to more functions in the AI system. In the near future, the system is expected to be able to tell the stage and predict the development of the disease. “We intend to gather data on the patients’ temperatures and blood tests as well to develop a multimodal learning model for our AI system,” said Xue.
Editor: Deng Jianguo
