Machines combating disease
Alejandro (Sasha) Vicente Grabovetsky, Co-founder of Avalon AI, discusses the ways in which machine learning is improving the rates of failed dementia clinical trials and improving the lives of those living with the disease.
Deep machine learning
The idea for Avalon AI came together when my Co-founder Olivier van den Biggelaar and I realised that we shared the same aim, which was to help defeat ageing. Following that, what immediately came to mind was dementia because it’s a disease that has not been successfully tackled yet. Lots of age related diseases like diabetes and cancer receive a lot of funding and are being heavily addressed, while dementia is under-funded partly due to failed clinical trials.
Very few dementia clinical trials have succeeded and we noticed that a lot of the past trials were targeting late-stage dementia, where a lot of brain damage had already occurred. We thought we needed to intervene at an earlier stage. This meant either at the stage where you have mild symptoms or even earlier, when you have initial brain damage, to try and stop the disease in its tracks there and then. The question was, how could we use our expertise in machine learning (e.g. deep learning) and MRI imaging to address this?
Early signs of dementia
Currently there are a couple of biomarkers to detect the progress of dementia, one is the size of the hippocampus, our brain’s memory chip, and the other is the size of the ventricles, which increase in size as our brain tissue degenerates. We wanted to take a fine grained approach, looking at changes in grey and white matter; as well as cerebrospinal fluid in the brain, and learn how they relate to the progress of mild cognitive dementia.
We do this by taking a 3D MRI image, aligning it to a template, and then using convolutional neural networks (CNNs). CNNs work a bit like the visual cortex of the person, with each layer of the network extracting simple features and recombining them into more complex ones. The neural network looks at both commonalities between people with dementia at similar stages and the differences between people at different stages. To analyse this information, we don’t need anyone’s private data, only their age, gender and perhaps their diagnosis, where this is available.
We need their gender because there are individual differences between males and females, such as the higher rate of women getting dementia than men, which is in part because they live longer.
Technology and dementia
Using deep neural networks to study dementia was not possible until the deep learning revolution three years ago, which allowed more complex models to be derived from data. This deep learning research has shown a lot of promise in classifying photographs, and has won many competitions (e.g. ImageNet), and has now started to be applied to medical research. However, most of the medical applications of deep learning have focused on supporting radiology decision for treatable diseases such as strokes and some types of cancer. This would be the first time that it gets applied to predicting the onset of dementia.
There are some challenges in looking at medical images, as it’s much easier to look at 2D photographs than 3D scans. This is due to the large image size of 3D scans and the need for the machine to process a lot more images in order for the neural networks to work. 3D medical images are far less available to use, because it’s difficult to gain access to them due to data protection and privacy concerns.
The other challenge is getting the neural networks to learn as fast as people. They learn slowly today because ‘learning’ requires solving a complex mathematical optimisation problem, which is difficult both conceptually and computationally.
How machine learning works
The algorithm doesn’t need to know if it’s a medical image or not, it just has to know it’s a three dimensional image. It knows that some of the images are of Alzheimer’s patients, some will be female or male, and also their age. It will then try to predict this data with a margin of error. The machine might say that this image is of a 50-year-old patient when in fact it’s wrong, so you then change the parameters in your model so that it gives a more correct answer.
This is where training and optimisation comes into play. The important thing is that the data is labelled in the way you want to predict it, meaning the computer corrects the neural network when it gets the data wrong.
We have had access to 80,000 scans and typically when we train a model, we go through 10,000 scans. While there have been promising results, it’s not ready yet to be used as a diagnostic tool. I believe we will get much better results when we gain access to 100,000 to a million scans; that will be the tipping point.
Our aim is to get people to live much longer and healthier lives. The more accurately we can diagnose health, the earlier we can intervene and reduce the damage diseases cause to the brain. Right now, we realise there is no cure for dementia and this is why we aren’t planning to immediately use our tool for diagnosis, but rather focus on making clinical trials cheaper and more successful.