AI assisting doctors treating macular degeneration

– I’m Adel Foda, I’m the Head
of Science at Silverpond. We have created automated
systems to solve problems in medicine, utilities, and
a range of other industries. And we’re largely focused
on computer vision, supported by our product HighLighter. It takes you through
the end-to-end process of developing a machine learning model, from creating a label annotated data set, making sure that it’s quality-assured, then finally monitoring
the model in production. – My name’s Devinder Chauhan, I’m a retinal specialist in Melbourne. I started a company called Macuject, which is a clinical decision
support software company, and used Silverpond to help me with this. When you need to do computer
vision or automated analysis using AI, it’s really
important to get annotation of the scans by experts,
to look at the scans and decide which bits are
abnormal, what they are, and outline those. – It can compare and
overlay their two answers, so that you can see the
points of difference. It gives you an opportunity to identify the more difficult cases that you may need to spend extra attention, to make sure that the
machine gets correct. – We could label scans
from wherever we were. I could do it in the evenings
or weekends from home in my spare time. That made life much, much easier. – The labelling that we had to do had a very specialised structure. Where we were looking for
layers inside the retinal scans, we were able to build a
plugin in to HighLighter, which encoded that structure by default, so that the labelers
wouldn’t have to spend time inputting that in to their annotations. And that was able to speed
up the labelling process. HighLighter gives you a complete history of the training data
that went in to training a particular instance of
a machine learning model that was deployed in production. We were able to monitor how
our labelers were progressing through the task, we were
able to plan their workloads by understanding the rate at
which they were able to label. To be able to compare it, see
it for ourselves visually, gain an understanding
of what they were doing, and then be able to provide feedback, so that ultimately we got the best outcome for the amount of times that
was spent with the labelers. – Our AI is already picking
up some tiny pockets of fluid that the doctors hadn’t seen on their normal looking through the scans. Even though I thought that
the product was gonna end up being one that was about saving time and introducing efficiencies, it’s actually introducing potentially a significant safety factor.

Leave a Reply

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