We tried our hand at developing a COVID-19 diagnostic application using Salesforce Einstein Image Classification
The COVID-19 pandemic (caused by the SARS-CoV-2 virus, colloquially known as the Coronavirus) has disrupted countries, communities and economies across the world and killed has killed (a conservatively estimated) 516,000 people. We are still learning the best ways diagnose and treat this virus but many countries have already relaxed their lock down orders (if they even implemented them to begin with), so it goes without saying that we’ll need to cooperate and collaborate to keep this virus under control in the coming months.
Technology can play a crucial role in helping us navigate this new COVID-19 world and could very well help us better manage future pandemics. Artificial Intelligence is being explored for its potential to help us diagnose and track infected patients more quickly, remotely and accurately.
Over the past few months, there has been a lot of discussion around the fact that chest x-rays are an effective way to detect COVID-19. Several research teams have announced the development of AI tools that they hope will be able to diagnose COVID-19 and other chest infections from x-rays alone (and be able to differentiate COVID-19 from other chest infections), removing at least some of the risk to diagnosticians and other patients from being repeatedly exposed to patients who may be COVID-19 positive. None of these have been made fully available to the public or medical community which makes it hard to assess their accuracy.
We decided to try our hand at building an AI app on top of Salesforce that would use Salesforce Einstein Image Classification to detect COVID-19 in patient x-rays. We have used Dr Joseph Paul Cohen’s (Postdoctoral Fellow at the University of Montreal) data to build this app.
The entire app is built on Salesforce, so it’s not only able to detect COVID-19 with its AI, but it is also able to perform all the out of box functionalities that Salesforce provides. This means it can be programmed to send emails to the proper authorities, patients, and doctors, utilise the data in dashboards to show the current infection rates based on region, and much more. The more data the app receives the more accurate its predictions will be.
Dataset Used:
The dataset consists of 60 COVID-19 positive patient x-rays and 70 x-rays from healthy patients. The most frequently used view is the posteroanterior (PA) view so we have used the COVID-19 PA view x-ray scans for our analysis.
You can watch a demo of the app below:
For the app, we have used an x-ray dataset which has 130 reference images out of which 46% are COVID positive and 54% are COVID negative. After we trained the dataset we got a unique MODEL id, which we used to predict the results. We stored the results within Salesforce and our dashboard shows you the graphical representation of it. We decided to group the testing data on the basis of months with the idea being that this will enable us to see how many tests are conducted each month and how many of those tests come back with a positive result.
Though this application, and the use of AI in diagnostic medicine are both in their infancy, it is clear that there is value in and a demand for this technology. Its potential to reduce medical costs, screening times, and the risk to healthcare professionals, patients, and the general public during a pandemic are significant. We are excited to see the role Salesforce and its partner community can play in developing these game changing programs and applications!
Disclaimer: This is just a Prototype/Proof of Concept. We appreciate that a larger number of cases is required before we can draw a meaningful conclusion from the data. We acknowledge that this testing would be more compelling if we could use the app to differentiate between COVID-19 pneumonia and pneumonia secondary to a non-COVID-19 pathogen, rather than identifying a positive COVID-19 patient against a patient who presents with a clear chest x-ray (who would be unlikely to be screened for the disease).
It is worth noting that an unknown but potentially significant percentage of the population can be both COVID-19 positive and asymptomatic. Those patients are unlikely to seek out medical care for the disease, or be approved for screening if they are not presenting any symptoms, and their chest x-rays would likely appear clear even if they were testing positive. At this stage, this application would not be able to either detect, account for, or track these asymptomatic COVID-19 positive patients.
If you have any thoughts or ideas regarding this app please contact us via the makepositive contact form. Thank You!
