Using Deep Learning for Anomaly Detection in Radiological Images

Deep learning is usually linked to big data, but in a medical context radiologists often need help in diagnosing rare conditions that have limited historical data. Can deep learning prove a useful tool in such cases?

Ioannis Katramados founded COSMONiO Health Lab in 2015 with a vision to develop intelligent medical technologies based on deep learning. COSMONiO's focus is on designing NOUS, an embedded deep neural network platform that aims to make high-accuracy predictions using significantly smaller training datasets. To achieve this the lab collaborates closely with UMC Groningen, one of the leading university hospitals in the Netherlands, where their research is currently focussing on performing automated anomaly detection on thoracic X-rays. 

I caught up with him ahead of his presentation on 'Anomaly Detection in Radiological Images Using Deep Learning' at the RE•WORK Deep Learning in Healthcare Summit on 7-8 April.

What are the key factors that have enabled recent advancements in deep learning?
The single most significant factor is the availability of low-cost GPUs. These allow us to train complex deep neural networks significantly faster than before. Since engineering often relies on trial and error, it makes sense that shorter training times enable deep learning engineers to perform more experiments even using a standard desktop PC. Alongside the contribution of GPUs, I should also acknowledge the importance of having great open-source libraries available from a wide range of academic institutes and companies.

How can deep learning methods help to diagnose rare conditions with limited historical data?
This is the main question that COSMONiO is trying to answer. There is no doubt that deep neural networks need a lot of data to become effective. However, when it comes to image classification we usually try to feed a vast amount of pre-labelled images to the neural network. But when big data is simply unavailable we have to invent other ways of training our algorithms. One way is to copy the human model, for example, when children are being taught to read and write. It is an interactive process involving the child and the teacher. In practice this means that we need better active learning systems, where user feedback plays an important role.

What are the main challenges to using a deep neural network platform to make high-accuracy predictions?
As with humans, the quality of predictions depends on the quality of the acquired knowledge. So one of the biggest challenges is the development of deep-learning frameworks that can handle noisy and inaccurate data more effectively, without relying on huge training databases. In this case, user feedback will be increasingly important, but of course we have to go beyond existing point-and-click feedback techniques that make this a laborious process.

What additional applications can COSMONiO NOUS be used for?
NOUS is an embedded deep learning platform that allows real-time training of neural networks. The hardware is based on cutting-edge NVIDIA technologies, while an incredible amount of innovation has gone into the software. We are currently developing a new deep learning engine that significantly reduces the computational overhead of training neural networks. This means that in the future we will be able to embed artificial brains in small devices that can learn throughout their lifetime. For non-embedded applications we will see a sharp increase in training speed and performance.

What developments can we expect to see in deep learning in the next 5 years?
1) E-brain hardware modules, 2) A deep-learning OS that writes its own software, 3) A revolutionary GUI for visualising deep neural networks. I am sure all three will happen sooner than we think.

Ioannis Katramados will be speaking at the RE•WORK Deep Learning in Healthcare Summit, in London on 7-8 April 2016. Other speakers include Alex Jaimes, CTO & Chief Scientist of AiCure; Brendan Frey, President & CEO of Deep Genomics; Ekaterina Volkova-Volkmar, Researcher at Bupa; Jeffrey de Fauw, Research Engineer at Google DeepMind and more.

Tickets are limited for this event, for more information and to register please visit the event page here.

Big Data Deep Learning Pattern Recognition Healthcare Deep Learning in Healthcare Summit MedTech Deep Learning Algorithms Computer Vision


    As Featured In


    Partners & Attendees

    Ibm watson health 3.001
    Rbc research.001
    Mit tech review.001
    Kd nuggets.001
    Maluuba 2017.001
    This website uses cookies to ensure you get the best experience. Learn more