To explain ‘deep learning’, let’s start with Artificial Intelligence (AI). In its simplest explanation, AI is anything that is trying to imitate a human process. ‘Deep learning’ (or machine learning) is a subset of AI, which is concerned with pattern recognition. For example, we can tell whether someone has a disease from an MRI scan because there is a pattern somewhere in the data from a population of patients. We're trying to find those patterns, and create AI models that can identify them automatically. 

Deep learning models do this using ‘features’. You have your input, which might be an MRI scan for example, and as the image is processed it gets turned into features. Humans can’t understand these, it’s the way the model learns from the data. Deep learning is – essentially – lots of these features in front of each other. The mathematics itself is fairly simple, but it’s used in long, deep neural networks which, given enough data, can learn incredibly complicated patterns. Interestingly, the exact same technology that powers driverless cars also powers deep learning for medical imaging. The tech telling you there's a pedestrian crossing the road is exactly the same as the tech telling you there's a lesion in the brain – their use depends on what data you give it. 

For these models to work well, you need to feed them a lot of data, example after example, until they learn the pattern. In a nutshell, that’s the deep learning component of AI.

How deep-learning AI is revolutionizing clinical data analytics

The clinical field is now realizing what’s possible using deep-learning AI – even though the concept of neural networks has been around for decades. About ten years ago it was discovered that you can take a graphics card – like those designed for gaming, which produces high-quality graphics and renders quickly – and apply the technology to optimizing a neural network. The computations are simple, you just have to do a lot of them. Once this was enabled, the revolution kicked off and people began applying these neural networks to many problems in a range disciplines, including the medical field. 

With a disease like Alzheimer's, where a certain part of the brain (in this case, the hippocampus in particular) will shrink over time, we need to be able to accurately measure changes in size over time. Neural networks applied to MRI scans make that possible, allowing us to determine what disease stage a patient is in and, potentially, whether a patient is responding to the drug that’s being trialed. 

While we’ve been answering these sorts of questions for years, we’re now starting to look at other kinds of questions like:

• Is this MRI scan fit for purpose?
• Is the scan good enough quality, do we need to get another?
• Does this subject have any safety concerns?

 Many drugs under trial carry a risk of side effects that need monitoring. For example, some might cause bleeds in the brain or a buildup of fluid – we want to develop algorithms that can automatically detect that, so we can intervene as soon as possible and, for example, change dosing to increase patient safety.

Benefits of CNNs in CNS clinical trials

Convolutional Neural Networks (CNN) have a number of benefits for CNS research and are probably the most widely used type of neural network when analyzing medical images. ‘Convolutional’ means there is a ‘window’ that runs along the image like a scanner, processing everything as it goes.

Greater flexibility

This window gives CNNs great flexibility, enabling a CNN to be quickly adapted to highly varied medical images. A neural network is named because it resembles neurons, mimicking the human brain. There are layers, and each layer is connected by a pathway of neurons. A standard neural network has one neuron connected to every part of the image, but the ‘window’ approach in a CNN means you end up with far fewer connections which hugely reduces the computational burden.  Therefore, CNNs are a highly cost effective solution.

Higher accuracy

CNNs are also very accurate when it comes to functions such as volume segmentation of the hippocampus and other brain structures. Typically, these deep neural networks outperform ‘classical’ non-AI approaches – something that we have really focused on demonstrating over the last 18 months.

How deep learning is evolving clinical trials

Often, medical imaging algorithms evolve from an idea first published in the computer science literature. However, when it comes to image segmentation, it’s been the other way around. The medical field was one of the first to make a type of CNN architecture far more accurate than was already out there. Since then, the technology has become ubiquitous across all fields. Wherever anyone wants to delineate something in an image, they're likely using a CNN first developed for medical image segmentation.

 Our core focus at IXICO over the last few years has been on volumetric segmentation. We’ve been asking things like ‘How big is the hippocampus?’ While this is not necessarily a simple question to answer, it is a structured one. There is a region in the brain we're interested in, we have the data, so we can answer it. Now, we’re starting to branch out and ask more ‘left-field’ questions about image quality and patient safety, with a focus on clinical trial enrichment.

Every time a patient drops out of a clinical study because of a safety finding or any other reason, it has an adverse impact on the trial sponsor’s ability to successfully show efficacy of their drug. Trial enrichment is about trying to find the best set of participants at the start. In Alzheimer's disease, patients’ symptoms will progress at various rates over a given timeframe, and a typical clinical trial lasts 2-5 years. In the trial enrichment phase, we want to identify candidates that are likely to respond well to the treatment. One way of doing that is to try and identify those patients that are likely to clinically decline over the period of the trial in the primary study endpoint. Finding patients that progress on this endpoint over the duration of the study increases chances to show a positive treatment effect. We’re working to get better at guessing – enabling drug development sponsors to gain critical efficiencies by optimizing sample size vis-a-vis statistical power.

Patient safety is also an evolving area when it comes to deep learning. For example, we’re looking at algorithms that have the ability to detect anomalies – even down to microbleeds which are typically only one millimeter across. Often, anomalies aren’t harmful, but they do need to be monitored and checked to make sure they don't develop into anything worse.

The future of deep learning AI in CNS research

The future of deep learning AI in clinical research is all about the data. The math has been there forever, the hardware has been there for about ten years, now it’s all about trying to understand how we can use the data. This is what will enable us to answer more valuable, clinically insightful questions around patient safety, disease progression, and image quality.

We’re at a very exciting stage in the evolution of deep learning AI in clinical research whereby we are using this existing data to try and formulate more experimental questions.

A unique approach to deep learning at IXICO

At IXICO, our key differentiators are our data and the experience with developing and deploying imaging biomarkers in CNS clinical trials. We have a team of highly trained, specialist image analysts spending a huge amount of time manually labelling and curating images as part of delivering our services.  The result is an ever expanding and diverse set of data points which we can feed into the neural networks, allowing them to learn by example and keep improving. These CNNs are now approaching human-level performance – which makes the technology not only accurate but also scalable.

Trial sponsors increasingly provide us access to their data for R&D purposes and, combined with the manual labelling we as part of the study, we now have access to a rich catalogue of exceptionally high-quality data. This puts us in a unique position to collaborate with consortia and universities who can review the data as part of their research in order to advance study into neurodegenerative diseases further.

We now have scientific publications showing our AI-driven methods for image analysis are performing better than non-AI approaches and manual analysis, and we have access to data that we can use to build better and better models. It’s an exciting time to be at IXICO and to be working in this field.

 

About the Author: Jack Weatheritt, PhD

Jack Weatheritt is a Senior Research Engineer at IXICO and has been with us for three years. He has a PhD in mechanical engineering and began his career working on AI development within the aerospace industry. After completing a postdoc position in Melbourne, Australia, he came back to the UK and joined IXICO’s technology team as a software developer. After a year, he made the transition to our science team and has been there ever since.

---