DeepMind’s New AI Diagnoses Eye Conditions


Dear Fellow Scholars, this is Two Minute Papers
with Károly Zsolnai-Fehér. In this video series, we often see how these
amazing new machine learning algorithms can make our lives easier, and fortunately, some
of them are also useful for serious medical applications. Specifically, medical imaging. Medical imaging is commonly used in most healthcare
systems where an image of a chosen set of organs and tissues is made for a doctor to
look at and decide whether medical intervention is required. The main issue is that the amount of diagnostic
images out there in the wild increases at a staggering pace, and it makes it more and
more infeasible for doctors to look at. But wait a minute, as more and more images
are created, this also means that we have more training data for machine learning algorithms,
so at the same time as human doctors get more and more swamped, the AI should get better
and better over time! These methods can process orders of magnitude
more of these images than humans, and after that, the final decision is put back into
the hands of the doctor, who can now focus more on the edge cases and prioritize which
patients should be seen immediately. This work from scientists at DeepMind was
trained on about 14 thousand optical coherence tomography scans, this is the OCT label you
see on the the left, these images are cross sections of the human retina. We first start our with this OCT scan, then,
a manual segmentation step follows, where a doctor marks up this image to show where
the most relevant parts, like the retinal fluids or the elevations of retinal pigments
are. Before we proceed, let’s stop here for a moment
and look at some images of how the network can learn from the doctors and reproduce these
segmentations by itself. Look at that! It’s almost pixel perfect! This looks like science fiction. Now that we have the segmentation map, it
is time to perform classification. This means that we look at this map and assign
a probability to each possible condition that may be present. Finally, based on these, a final verdict is
made whether the patient needs to be urgently seen, or just a routine check, or perhaps
no check is required. The algorithm also learns this classification
step and creates these verdicts itself. And of course, the question naturally arises:
how accurate is this? Well, let’s look at the confusion matrices! A confusion matrix shows us how many of the
urgent cases were correctly classified as urgent, and how often it was misclassified
as something else and what the something else was. The same analysis is performed to all other
classes. Here is how the retina specialist doctors
did, and here is how the AI did. I’ll leave it there for a few seconds for
you to inspect it. Really good! Here is also a different way of aggregating
this data – the algorithm did significantly better than all of the optometrists and matched
the performance of the number one retina specialist. I wouldn’t believe any of these results if
I didn’t see these reports with my own eyes in the paper. An additional advantage of this technique
is that it works on different kinds of imaging devices and it is among the first methods
that works with 3D data. Another plus that I really liked is that this
was developed as a close collaboration with a top tier eye hospital in London to make
sure that the results are as practical as possible. The paper contains a ton of more information,
so make sure to have a look! This was a herculean effort from the side
of DeepMind, and the results are truly staggering. What a time to be alive! Thanks for watching and for your generous
support, and I’ll see you next time!

3 Comments

Add a Comment

Your email address will not be published. Required fields are marked *