By 12th June 2020, COVID-19 had infected over 7.6M individuals and resulted in more than 420,000 deaths. High resolution computed tomography (HRCT) of the lung is the front-line tool for monitoring COVID-19 appearances [1-4].
HRCT facilitates accurate assessment of disease severity, progression or response to treatment. However, the sensitivity of HRCT for detection of lung lesions can be as low as 70% for experienced (5) and 51% for less experienced radiologists (6) highlighting the urgent need for rapid upskilling of clinicians regardless of their previous experience.
Solution
We present an educational solution – CovED based on a platform that has revolutionised other radiologic domains. From the beginning of April-2020 the prototype has been made available free to every clinician across Australia and has overnight been adopted on every continent. Our technology is based on effective intelligent education, which improves radiologic detection of disease by 34% (7). It is web-based, allowing clinicians located anywhere to diagnose de-identified lung CT images with known truth. Our algorithms rapidly calculate the clinician’s diagnostic ability and can instantly identify any errors made.
PhD, MSc. IT, BSc. Applied Physics,
Co-Founder DetectED-X
We are developing CovED as an online platform. It’s based on our previous breast and lung cancer models [7] using REACT programming language and MySQL database. The presentation interface is the same as that used clinically. Individuals will access the software at the DetectED-X website using the Chrome internet browser. The register takes a matter of seconds to access and complete a brief questionnaire. Once completed, the reader can choose one of several CovED modules. Each presents unique patient cases demonstrating a series of axial lung CT slices. Any imaging platform (e.g. MAC or PC) can be used.
Specifically, some CT cases will demonstrate COVID-19 appearances, some will not. Each individual will diagnose the cases using drop-down menus and post-processing options typically used in clinical practice.
As shown in Figure 1, each individual will be asked to comment on whether the case contains the appearances of:
- ground glass opacity;
- crazy paving/mosaic attenuation and/or consolidation; and
- on the location of any perceived appearance.
After all these appearances have been considered, the user must give a confidence score from 0-5 on whether the case is COVID-19 positive or not. Observers can correct any previous decision. The process takes from 30 minutes to 1 hour to complete.
Truth files for each patient case will be developed from expert consensus of senior respiratory radiologists who judge each clinicians’ interaction with these cases.
Using this truth, all individual interactions with the images are instantly analysed.
Once all cases are diagnosed, clinicians are immediately presented with performance values for sensitivity and specificity calculated using the 0-5 scale described above. Receiver operating characteristic (ROC) values are defined from the area under the curve produced by plotting sensitivity vs 1-specificity and the confidence scores given to each case.
This is followed by a case by case review where algorithms compare individual interactions for each image to that of the expert consensus. At this stage, an expert synoptic report is available for each case, using the Radiological Society of North America categories for reporting COVID-19 appearances. They are:
- typical;
- indeterminate;
- atypical; and
- negative.
For example, a typical synoptic report provided by our experts would be:
Extensive bilateral ground glass opacities and consolidation with a peripheral and posterior predominance. Small bilateral effusions. The imaging appearances are typical of COVID-19 and consistent with severe disease.
Finally, COVED provides a certificate of completion with Continuing Medical Education scores.
On-going work
We are now working closely with our key collaborators to enable a rapid development in sophistication, security and effectiveness of the CovED platform. Collectively, we will be able to use the CovED tool to assess COVID-19 disease severity, its progression and its response to treatment.
Over the next year we will deliver:
- A free on-line platform that will build on the current prototype to enable:
- Multiple test sets of images from a variety of countries and vendors;
- Tailored test sets to match individual clinician capabilities;
- AI-based allocations of specific image types to match individual clinician weaknesses;
- AI companion to support clinical decision-making with difficult cases.
- Accreditation by the Accreditation Council of Continuing Medical Education (ACCME).
Who’s involved in the COVED project?
DetectED-XÂ Â (a University of Sydney Start Up set up by Professor Patrick Brennan, Professor Mary Rickard and Dr Moe Suleiman)Â leads the CovED project consortium involving:
- the COVID-19 Image Biobank (led by Professor Stuart Grieve and a large consortium of University of Sydney clinical experts;
- GE Healthcare Volpara;
- World Continuing Education Alliance (WCEA); and
- Amazon
This project will ensure widespread and robust adoption of CovED.
References
- Tao AI, et al. Radiology Supplement 2020 Special Focus: COVID-19 https://doi.org/10.1148/radiol.2020200642
- Kanne JP. Radiology Supplement 2020 Special Focus: COVID-19 https://doi.org/10.1148/radiol.2020200241
- Bernheim A, et al. Radiology Supplement 2020 Special Focus: COVID-19 https://doi.org/10.1148/radiol.2020200463
- Ming-Yen Ng, et al. Radiology Supplement 2020 Special Focus: COVID-19 https://doi.org/10.1148/ryct.2020200034
- Al Mohammad B, et al. Clin Radiol. 2019 Jan;74(1):67-75.
- Tsim S, et al. Lung Cancer. 2017 Jan;103:38-43.
- Suleiman, W. I. et al. J Med Imaging Radiat Oncol, 2016 60(3), 352-358. doi: 10.1111/1754-9485.12461.
P Brennan PhD, M. Suleiman PhD and T Davies,
12 June 2020
