Date: 2nd March 2020
Article in brief:
The clinical manifestations of COVID-19 pneumonia are complicated and can be characterised as fever, cough, myalgia, headache, and the onset of gastrointestinal symptoms. Whilst rapid detection kits for diagnosing COVID-19 are becoming more readily available, computed tomography (CT) scans are still the most efficient way to diagnose and evaluate the severity of the pneumonia.
Now a team from Renmin Hospital of Wuhan University, China, led by Honggang Yu, has constructed and validated a system based on deep learning for the identification of COVID-19 pneumonia on high resolution CT. The model has comparable performance with expert radiologists but is considerably faster.
In the early days of the COVID-19 outbreak the team observed a concerning bottleneck with traditional diagnosis, as thousands of patients queued for CT scans. With many patients requiring in depth, slow scans, radiologists were overloaded leading to massive delays in diagnosis, isolation of patients, and subsequent patient treatment and prognosis. The team – no strangers to AI – had previously used it to detect minor lesions in gastrointestinal endoscopy. They therefore reasoned that, by applying deep learning (DL) to detect tiny features during image analysis, it could provide a highly valuable diagnostic tool for radiologists and doctors, and contribute to the control of disease.
In the pre-print paper published on medRxiv, the team have developed, and validated a deep learning model.
- Over 46,000 CT images, from 106 admitted patients, including 51 patients with confirmed COVID-19 pneumonia, were retrospectively collected and processed and used for model development and training.
- A further 27 patients added to the study were evaluated and compared against the efficiency of radiologists using the deep learning algorithm.
- These 27 patients were also diagnosed using respiratory secretions samples using a US National Drug Administration approved COVID-19 nucleic acid detection kit to confirm the AI-based diagnosis.
- The model achieved a per-patient sensitivity of 100%, and diagnosed 16 patients as positive, as did the human equivalent.
- Overall, the model (measured by several parameters) showed comparable performance to the expert radiologist.
- With the assistance of the DL model the reading time of the radiologists was decrease by 65%.
Conclusion:
We are currently seeing a massive drive around the world to slow down the spread of COVID-19. By trying to contain the disease, it gives us time to find effective treatments or cures, and to allow the healthcare systems time to prepare and not become overwhelmed by the volume of patients.
We have already seen the power that AI can bring to the epidemiology of COVID-19, from predicting the spread of infections, using data to understand the nature of the virus, to aiding complex decision making. AI-driven vaccines and drug discovery could vastly increase the rate of effective treatment development and implementation. Now, here, the speed of diagnosis, will also hopefully aid the containment, and allow the rapid treatment, and isolation of those infected; a crucial factor in patient prognosis.
Chen, J., L. Wu, J. Zhang, L. Zhang, D. Gong, Y. Zhao, S. Hu, Y. Wang, X. Hu, B. Zheng, K. Zhang, H. Wu, Z. Dong, Y. Xu, Y. Zhu, X. Chen, L. Yu and H. Yu (2020). “Deep learning-based model for detecting 2019 novel coronavirus pneumonia on high-resolution computed tomography: a prospective study.” medRxiv: 2020.2002.2025.20021568.
https://doi.org/10.1101/2020.02.25.20021568
