Let’s talk.
Reach us anytime via weguide@support.com
Let’s talk about
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
Back to Blog

Using Neural Networks in Healthcare

Product Engineer, Cliff, outlines how neural networks are currently being used in healthcare to help analyse retinal OCT images

Cliff, one of our Product Engineers, routinely works with coders, clinicians and clients to help create beautiful solutions for some of healthcare's greatest challenges.

But most recently, he's been kicking goals with RMIT STEM College, who approached us to do a case study on Cliff’s work integrating Machine Learning in retinal scan analysis for Dr Devinder Chauhan's Macuject.

Macuject is a Melbourne-based Medtech startup that aims to prevent blindness in 2.5 million people a year by guiding doctors and patients to personalised treatment strategies throughout their life-long eye injection journey.

Cliff's work will be part of a STEM-related industries course, shown to over 6,000 students in the near future!

Thanks go to the following team making this all possible: Professor Aleks Subic, Professor Angela Carbone, Dr Suneeti Rekhari, Simon Gibson, Yolanda Rios and RMIT STEM College.

Video Transcript

My name is Cliff Warren, and I'm a product engineer and a technical leader at Curve Tomorrow, a digital health tech company that has been working in the industry for over 10 years and have been partnering with a lot of research institutes and hospitals to try and help get their research into the hands of our patients and customers.

Curve Tomorrow was originally engaged by a retinal specialist named Dr. Devinder Chauhan, three years ago, to develop a clinical decision support tool. Originally, he had noticed that there was a great variation in the type of care that was provided in a number of different retinal clinics. He saw an opportunity to streamline some of the processes within his organisations. So, he engaged us to try and create a clinic decision support tool that would allow for efficiencies within the clinic and an equal level of care.

Through interviewing the clinicians and the nurses and the patients, and over the course of developing that product, we discovered that one of the major pain points was in reading the OCT scans that patients were receiving in order to determine if they had macular degeneration. We reviewed this problem and realised that it would be a great case study to use some machine learning, and particularly image processing in combination with it, to do some segmentation of those OCTs to provide some insights for the clinicians to enable them to more easily and readily identify pathologies within the eye.

In doing interviews with clinicians, we discovered that the level of expertise required to correctly analyse the OCTs meant that a lot of senior doctors needed to be part of the decision-making process. One of the obvious advantages of having this safety guard with the with the machine learning inferencing is it allows junior clinicians to be able to provide that same level of care.

In working with the clinics, we noticed that the average time to see a patient - including reviewing the OCTs - was sometimes up to 10 minutes. And with that, the quality of care relied on that senior clinician to provide insights and allow for the best level of care. With the AWS (Amazon Web Service) machine learning technology in place, we allow those same clinicians to spend more time with the patients and have more certainty in the decisions that they're making, allowing for nuanced care without changing the level of safety and the care that the patient receives.

The engagement with Devinder and his clinical team was absolutely critical to the success of the project. Without having the buy-in of the key opinion leaders, the users that will eventually use the product, and getting their expertise and insight into how you can use the machine learning itself is definitely the first step to any successful project. We worked with their team, and they spent many, many hours labelling the existing OCTs in order to create a robust training set that we're able to then use to feed into training the machine learning model.

Over the course of a number of months, we've worked with people from Devinder’s industry himself and a number of other clinics in order to validate that, which is, once again, probably the biggest hurdle that you might come across when trying to solve problems with machine learning is making sure that people in the industry want that problem solved.

The process involves sending an OCT to the machine learning inference platform on AWS, it will then take that OCT and slice it into the vertical slices and then analyse each of those images and provide us with an image segmentation that gives us an inference of the volumes of fluid within the eye. The fluid that is present in the eye is one of the key indicators of a loss of vision. What Macuject aims to do is to identify them early and provide the best level of treatment in order to reduce the fluid.

When a clinician sees a patient, they will record an OCT of both of the patient's eyes. That OCT machine will pass that OCT image up to our platform in AWS, which will segment the image and return to us whether or not any fluid has been detected in the retina. The clinician has the ability to then open that via our platform and compare side by side what the original OCT would look like, and then what our inference is providing. The impact that this has on patients is that they need to have less visits from a very expensive doctor and a better outlook and better vision.

Within the clinical setting, the OCT would be viewed by a senior ophthalmologist, and they would review the scans to determine whether or not there was fluid present in the eye. There are a number of different types of fluid that can be present and the machine learning platform has the ability to create a distinction between the interactional fluid and subretinal fluid, which are two different key indicators of potential loss of vision.

Across the globe, there's a number of research institutions that are starting to work together to provide those very, very difficult-to-come-by datasets that are pre-labelled by experts, and are really leveraging that sort of global community to provide some more insights into the data that is available from different backgrounds and different countries.

So, one of the biggest problems is that this is just one of a number of inputs into the clinical decision support. So, this is aiding and saving a bunch of time and giving a competitive advantage, because you can detect things earlier. Another one of the major speed humps or hurdles, perhaps with any machine learning application is understanding the regulatory parts required. So, the FDA, which is the Federal Drug Administration in the US and the TGA, which is the Therapeutic Goods, Australia, are both the two regulatory paths that we've needed to seek for this particular product.

I think the future holds a great opportunity for machine learning with a number of different pathologies, even within just the retinal space alone; I know that there are a number of different research groups that are currently working on diabetic retinopathy and other associated eye diseases to really help curb the global blindness issue that we have with particularly in the ageing population. On a larger scale, machine learning has the opportunity to really take away a lot of the difficult or dangerous or just dull aspects of a lot of healthcare and a lot of other industries as well.

Share the article