Diagnosing COVID-19 using AI-based medical image analysis

The COVID-19 disease, caused by the SARS-CoV-2 virus, colloquially known as the Coronavirus, is disrupting large parts of the world. We know little about the best way to get to a diagnosis and about what prognoses are implied, let alone what treatment works best in specific situations. We all need to join forces to get this virus under control. And by all, we mean both human and machines. Technology is crucial to get us through this and a special role for artificial intelligence (AI) is to be expected. How can AI support us in this fight against COVID-19? What means can AI provide to help us diagnose patients quickly and accurately?

In this article, we discuss the current methods for diagnosis of COVID-19 and investigate the possible role of artificial intelligence in radiology.

What does the current diagnostic pathway for COVID-19 look like?

If there is a suspicion of COVID-19 combined with other factors such as that the patient has been in suspected territory, has been in touch with a person already diagnosed with COVID-19, or requires hospitalization, an official diagnosis needs to be made. What is currently the most common way of diagnosing COVID-19? What other options do we have and what role can we expect for radiology in this process?

Laboratory testing for COVID-19 diagnosis
There are several ways to run a laboratory test on a patient’s specimen, by far the most common one being RT-PCR. Briefly explained, this works in the following way. A sample is collected from the patient, usually with a nasopharyngeal or throat swab. Through a chemical process, the patient’s RNA is collected from this sample and a specific enzyme (a so called “reverse transcriptase” enzyme, hence RT-PCA) is added to turn the RNA into two-stranded DNA. Subsequently, one adds nucleotides, an enzyme causing the DNA to multiply, and short synthesized DNA fragments called “primers.” These primers, combined with fluorescent dye, are able to signal when the sample contains viral DNA, i.e. the patient tests positive for COVID-19.1

Currently, this is the standard way to test for COVID-19. The specificity of this test is good; however, the sensitivity can vary greatly depending on the country. In the short time that the virus emerged, each country developed different tests based on the same principle, leading to sensitivity ranges varying from 60-70% to 95-97%. Hence, some geographies are coping with a significant number of false negatives.2

Other options for lab testing include “molecular point-of-care” tests (a mostly automated type of test, which can be performed by a frontline healthcare worker - results are available within 30 minutes instead of days), paper-based tests (similar to the way a pregnancy test works, a paper strip contains antibodies that will bind to viral proteins in a sputum sample),3 or testing for antibodies in the blood (not a very relevant option, as the antibodies can only be measured after being infected for days or even weeks). Unfortunately, none of them has a version ready for clinical use that provides fast results and good accuracy.

Medical imaging for COVID-19 diagnosis

Next, there are medical imaging-based methods of which CT is the most talked about contender for fast and accurate diagnosis of COVID-19. Sensitivity of a CT-based COVID-19 diagnosis is reported to be significantly better than RT-PCR can offer (80-90%), however, specificity is on the low side (60-70%).4 How to recognize COVID-19 on a chest CT? Described findings include ground-glass opacities, crazy paving appearance, air space consolidation, bronchovascular thickening and traction bronchiectasis.2 For the well informed reader, yes, these symptoms are very similar to regular pneumonia. Bai et al. investigated this fact and found high specificity, but moderate sensitivity for radiologists that were asked to distinguish COVID-19 from viral pneumonia.5

Additionally, one should bear in mind that contamination of scanners is a real issue.6 Radiology personnel has to be extremely diligent when it comes to cleaning scanners in between patients. In this light, chest X-rays can be an alternative as machines are easier to clean. Reporting on using X-ray for diagnosing COVID-19 has been ambivalent so far though. Hospitals in Spain mention the technique is a default element of their diagnostic pathway, while other sources describe X-ray as an insensitive test.2,7 Future research shall have to provide more clarity.

Peng et al. describe a strong correlation between ultrasound (US) and CT findings making US a promising technique as it is superior to chest X-ray when it comes to pneumonia exams, it is easy to use at the bedside, and it does not use ionizing radiation.8–10 However, US requires closer contact between the physician and the patient, which may increase contamination risks for the staff - something that should not be neglected. As little research has been done so far - also for US-based diagnosis of COVID-19 - future results shall have to provide more insight.

Furthermore, PET-CT is mentioned as a candidate for diagnosing COVID-19. Besides this technique being relatively time consuming, so far, research has been highly quantitative in nature and suggests that PET-CT can only add value if complex cases require additional information for a differential diagnosis.11,12

Have leading authorities already published any guidelines on the matter? Yes, they have. During the first half of March 2020, the ACR published guidelines concerning medical imaging for COVID-19 diagnosis. It is clear their message advises against the usage of medical imaging as a first-line test for diagnosis, for the moment. However, CT and X-ray are mentioned as optional methods for specific cases.13

AI to support medical imaging for COVID-19 diagnosis?

It has already been widely recognized that AI can potentially have a substantial role in streamlining and accelerating the diagnosis of COVID-19 patients. Several open dataset initiatives have been set up over the past weeks to enable the AI-community to develop and test methods that could contribute to countering the Corona pandemic.14–16

How can AI help analyze CT images of COVID-19 patients?

AI-based CT assessment is seen as one of the promising techniques that might lift some of the heavy weight of the physicians’ shoulders.17 Rapidly, research groups are demonstrating deep learning-based proof of principle18 or building prototype AI-algorithms that can help detect COVID-19 on chest CT scans. For example, the work of Gozes et al., published on March 10, 2020, reports a sensitivity of 98.2%, but also an impressive specificity of 92.2% for their deep learning-based thoracic CT algorithm.19 While there are initiatives focusing on the segmentation and quantification of lung infection regions,20 many target mainly the findings related to pneumonia, which is often present in COVID-19 patients.21 This is also the case for a range of initiatives set up by companies. For example, the software made available by Infervision is based on their CT pneumonia tool.23 As pneumonia findings are strongly related to COVID-19, such tools will come in handy for sure, but how to differentiate between COVID-19 and pneumonia with different causes? We are not the only ones to ask this question. Alibaba, for example, claims to be able to differentiate between COVID-19-based pneumonia and other pneumonia cases with an accuracy of 96%22 and a very fresh paper from Li et al. on distinguishing COVID-19 from community acquired pneumonia based on chest CT claims a sensitivity and specificity of 90% and 96% respectively, for detecting COVID-19. Bear in mind, though, that researchers used RT-PCR as a ground truth for which accuracy numbers are not mentioned in the article.24

Additionally, some companies have already started initiatives to develop algorithms for dedicated COVID-19 detection on CT images using scans of people diagnosed with the virus. Their aim is to get these solutions out there as fast as possible and offer them for research purposes.25,26

What about X-ray and ultrasound?

Other imaging modalities have received less attention from the AI-community so far. X-ray, while easier to obtain in most clinical settings, has only been sparingly investigated, and there have been no reports on US-based artificial intelligence methods so far.27–29

New call-to-action

AI-based medical imaging of COVID-19 in the future

The aforementioned arguments suggest that a combination of RT-PCR with chest CT exams will lead to a diagnostic protocol with both high sensitivity and high specificity. However, in a realistic clinical context of the current COVID-19 pandemic, speed, availability, and ease of application are of utmost importance; therefore, even a combination of both methods does not seem ideal. Will AI be the magical solution in this situation? As described above, application of AI methods to CT scans are being actively investigated and show promising results. However, AI will not make CT available at bedside, nor will it speed up the RT-PCR process.

US is definitely offering many benefits; a point-of-care method, no ionizing radiation, and more widely available than CT. Investigating the ability of this modality to support an (early) COVID-19 diagnosis would therefore most certainly be interesting. Especially if (to be developed) AI methods can support the less experienced sonographers.

All in all, the current situation requires fast, widely accessible diagnostic tools which would preferably have been available yesterday. Taking into consideration that developing certified AI radiology software usually does not happen within a week, medical imaging AI does not seem ideally suited for the immediate support that is much needed in the current situation. However, it might provide valuable preparation for future cases.

 

Bibliography

  1. Molteni, M. & Rogers, A. Everything You Need to Know About Coronavirus Testing. (2020). Available at: https://www.wired.com/story/everything-you-need-to-know-about-coronavirus-testing/. (Accessed: 19th March 2020)
  2. Moore, C. M. & Bell, D. J. COVID-19. Available at: https://radiopaedia.org/articles/covid-19-2. (Accessed: 19th March 2020)
  3. Trafton, A., Chu, J. & Chandler, D. L. Covid-19 diagnostic based on MIT technology might be tested on patient samples soon. (2020). Available at: http://news.mit.edu/2020/covid-19-diagnostic-test-prevention-0312. (Accessed: 19th March 2020)
  4. Hou, H. et al. Correlation of Chest CT and RT-PCR Testing in Coronavirus Disease 2019 (COVID-19) in China: A Report of 1014 Cases Tao. 2019, (2019).
  5. Bai, H. X. et al. Performance of radiologists in differentiating COVID-19 from viral pneumonia on chest CT. Press Radiol. (2020).
  6. Mossa-Basha, M. et al. Radiology Department Preparedness for COVID-19: Radiology Scientific Expert Panel Mahmud. Press Radiol. (2020).
  7. Chen, N. et al. Epidemiological and clinical characteristics of 99 cases of 2019 novel coronavirus pneumonia in Wuhan , China : a descriptive study. Lancet 395, 507–513 (2020).
  8. Peng, Q. Y. et al. Findings of lung ultrasonography of novel corona virus pneumonia during the 2019 – 2020 epidemic. Intensive Care Med. 6–7 (2020). doi:10.1007/s00134-020-05996-6
  9. Stempniak, M. Radiologists ‘strongly recommend’ bedside ultrasound for COVID-19 pneumonia diagnosis. (2020). Available at: https://www.radiologybusiness.com/topics/care-delivery/radiologists-ultrasound-covid-19-chest-x-ray-coronavirus. (Accessed: 19th March 2020)
  10. Poggiali, E. et al. Can Lung US Help Critical Care Clinicians in the Early Diagnosis of Novel Coronavirus (COVID-19) Pneumonia? Press Radiol. 9–11 (2020). doi:10.1148/radiol.2020200370
  11. Sijuan, Z. & Xiaohua, Z. FDG PET/CT of COVID-19. Press Images Radiol. (2019). doi:10.1148/radiol.2020200527.
  12. Qin, C., Liu, F., Yen, T. & Lan, X. F-FDG PET / CT findings of COVID-19 : a series of four highly suspected cases. (2020).
  13. Palmer, W. J. ACR Releases CT and Chest X-ray Guidance Amid COVID-19 Pandemic. (2020). Available at: https://www.diagnosticimaging.com/coronavirus/acr-releases-ct-and-chest-x-ray-guidance-amid-covid-19-pandemic. (Accessed: 19th March 2020)
  14. Hatmaker, T. With launch of COVID-19 data hub, the White House issues a ‘call to action’ for AI researchers. (2020). Available at: https://techcrunch.com/2020/03/16/coronavirus-machine-learning-cord-19-chan-zuckerberg-ostp/guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_ sig=AQAAAFXM3qnCspbtPxBBuSEn8Wye4PMTe9MTwMJDQcJ9XKXpS9HopZIpEEP01YMTWbuY_a3iOFH16wy6Ah_jG3SJQnWDpRVpVHJ69sErLgEROwTtIxZIUEP2QC6JM7XMhuXGdAw8bV_oDpbgpw2C1XJQwhMmJWZpzlx5RCZ-TupfG0Hs. (Accessed: 19th March 2020)
  15. Novel Corona Virus 2019 Dataset. (2020). Available at: https://www.kaggle.com/sudalairajkumar/novel-corona-virus-2019-dataset. (Accessed: 19th March 2020)
  16. covid-chestxray-dataset. (2020). Available at: https://github.com/ieee8023/covid-chestxray-dataset. (Accessed: 19th March 2020)
  17. Mccall, B. News COVID-19 and artificial intelligence : protecting health-care workers and curbing the spread. Lancet 2019, 2019–2020 (2020).
  18. Wang, S. et al. A deep learning algorithm using CT images to screen for Corona Virus Disease (COVID-19) Shuai. medRxiv 1–26 (2020). doi:doi.org/10.1101/2020.02.14.20023028
  19. Gozes, O. et al. Rapid AI Development Cycle for the Coronavirus ( COVID-19 ) Pandemic : Initial Results for Automated Detection & Patient Monitoring using Deep Learning CT Image Analysis Article Type : Authors : Summary Statement : Key Results : List of abbreviations : Submitt. to Radiol. Artif. Intell. (2020).
  20. Shan, F. et al. Lung Infection Quantification of COVID-19 in CT Images with Deep Learning Author. (2020).
  21. Xu, X. et al. Deep Learning System to Screen Coronavirus Disease 2019 Pneumonia. 1–29 (2019).
  22. Kharpal, A. China’s giants from Alibaba to Tencent ramp up health tech efforts to battle coronavirus. (2020). Available at: https://www.cnbc.com/2020/03/04/coronavirus-china-alibaba-tencent-baidu-boost-health-tech-efforts.html. (Accessed: 19th March 2020)
  23. Infervision Products. (2020). Available at: https://global.infervision.com/product/5/. (Accessed: 19th March 2020)
  24. Li, L. et al. Artificial Intelligence Distinguishes COVID-19 from Community Acquired Pneumonia on Chest CT. Press Radiol. (2020).
  25. Alberich-Bayarri, Á. Imaging, AI and Radiomics to understand and fight Coronavirus COVID-19. (2020). Available at: https://quibim.com/2020/02/14/imaging-ai-and-radiomics-to-understand-and-fight-coronavirus-covid-19/. (Accessed: 19th March 2020)
  26. Aidence. Aidence is working on an AI solution for COVID-19 on chest CT. (2020). Available at: https://www.linkedin.com/posts/aidence_johns-hopkins-coronavirus-resource-center-activity-6645980918089596928-8Rt3/. (Accessed: 19th March 2020)
  27. San Joaquin, A. Using Deep Learning to detect Pneumonia caused by NCOV-19 from X-Ray Images. (2020). Available at: https://towardsdatascience.com/using-deep-learning-to-detect-ncov-19-from-x-ray-images-1a89701d1acd. (Accessed: 19th March 2020)
  28. Rosebrock, A. Detecting COVID-19 in X-ray images with Keras, TensorFlow, and Deep Learning. (2020). Available at: https://www.pyimagesearch.com/2020/03/16/detecting-covid-19-in-x-ray-images-with-keras-tensorflow-and-deep-learning/. (Accessed: 19th March 2020)
  29. Alberich-Bayarri, Á. QUIBIM R&D team has just launched an updated version of the #AI model for #covid19. (2020). Available at: https://www.linkedin.com/posts/angelalberich_quibimprecision-quantitative-imaging-biomarkers-activity-6645740427137941504-q1j2. (Accessed: 19th March 2020)