New applications of machine learning technology and AI frameworks have enabled researchers and developers to quickly develop new systems meant to detect the “CCP virus” as it makes its way through the human body. In the following image provided by the IEEE Spectrum organization — one of these new technologies is a potential “CCP virus detection” tool that uses CT scans from patients to automatically measure and quantify the spread of the virus on the lungs.

In order to build these types of machine learning tools — the idea follows a pattern similar to this:

  1. “Measure” by gaining access both “negative” and “positive” image samples — where “positive” samples indicate a set of infected lungs, and “negative” means a set of healthy lungs.
  2. “Train” with data — by giving the system as many “annotated images” (typically, this is done through the use of “bounding boxes” where a human [or machine] outlines the location of the infection on the image) of infected lungs as possible. These “positive” samples are then marked with a “1” by the system. Healthy lungs are “negative” and therefore would register a “0” — though this example is greatly simplified.
  3. “Validate” by testing a set of known infected lungs against the new machine learning model. This is done in order to measure statistics like the “D-Prime” metric — meant to detect how accurate the model is in the real world against “known knowns”. When the “positive” samples are tested by the new machine learning model, the model is able to self-score the level of accuracy on its own. Typically, a “higher score” (closer to 1.0) means a more accurate model, assuming that the validation framework has been implemented properly.
  4. “Deploy” the model by allowing medical workers to simply take a CT scan, upload it to a remote (perhaps cloud) server, and then receive feedback in near real-time — hopefully alleviating some of the requirement for skilled doctors to review patients scan results manually.

To greatly simplify this practice — the system is “trained” to automatically detect “features” (like the presence of the virus on the lungs) in images that are uploaded by remote medical staff responsible for taking CT scans.

One major application of this type of research is to reduce pressure on already stressed doctors responsible for diagnosing potentially many cases simultaneously who would otherwise be left to manually review these images on a case-by-case basis.

For more information: