Whenever one is using Artificial Intelligence in high-stakes situations like medical diagnosis, one always worries about the false negatives i.e. when one tells a sick patient that they are healthy and the less critical case of a false positive, i.e. when one tell a healthy patient they are sick.
Thus one wants as accurate predictions as possible, but not only that. One wants to accurately estimate the uncertainty in the predictions, such that there is a way for the model to communicate the need for further scans and/or involve a medical professional.
In addition to that, the retinal scans need to be processed in high resolution to detect all the possible signs of pathology, a lot higher than normal machine learning algorithms are designed to work on.
We utilized Bayesian Deep Probabilistic Programming, to construct a model that was capable of determining if the scan were positive or negative for diabetic retinopathy. Since we used Bayesian Inference to do this, we were also able to provide the corresponding uncertainty estimation, thus providing much-needed information in the decision-making process. Such one could flag scans where the classification was uncertain and a medical professional should be consulted.
By providing a medical scan that minimize the need for a medical professional, one makes the scanning technology more accessible, faster, and cheaper. At the same time, it ensures consistent high-quality interpretations of the scans removing the variability of having humans interpret the scans.