When deep learning meets image registration

What if, instead of explaining to a computer how to solve a problem, you let it learn by itself? Within the area of machine learning, deep learning methods are revolutionising artificial intelligence and computer vision. They offer tremendous opportunities for machines that learn to perform human activities like listening and seeing. For instance, when it comes to fundamental vision tasks, like image classification or object recognition, a particular class of deep-learning algorithms has proved to outperform all existent strategies. Dr. Enzo Ferrante now aims to understand how we can use it to solve the task of image registration – the process of aligning images of a same object coming from different devices, moments or viewpoints. His objective is to allow for faster and more accurate registration methods and toolboxes, leading to more reliable image-based understanding studies and faster decision-making processes.

The production of more and more image data, whether from satellites, mobile phones or medical imaging, offers tremendous opportunities in the fields of environmental and life risks understanding. « For example, the damage produced by a natural hazard could be measured by aligning satellite images captured before and after the event and drawing comparisons between them », Dr. Enzo Ferrante explains. « Medical images can also be analysed by means of image registration, comparing them and trying to find patterns between patients with and without a certain pathology ». « Image registration is a fundamental problem in computer vision. Numerous research exists on the topic but most of it aims to explicilty model and explain to the computer how to align the images properly ». « Now, with the advent of the big data era and deep-learning methods, we know that a different approach exists: letting the computer learn by itself ». « In the field of computer vision, this new strategy has proved easier and more accurate », Dr. Enzo Ferrante reports, adding that this is a change of paradigm.

A computer vision technique inspired by the biological brain

The class of deep-learning models Dr. Enzo Ferrante aims to apply to image registration are called deep convolutional neural networks (CNNs). As the term “neural network” suggests, these computational models are inspired by biological brains and can perform complex feats of intelligence, while using little pre-processing compared to other machine learning algorithms. Still, to reduce the amount of image data necessary for training such a network, Dr. Enzo Ferrante intends to incorporate some prior-knowledge into the CNN-based image registration model. His idea is to go beyond simple visual observations and introduce context coming from prior knowledge such as anatomy.

Images can contain hidden information crucial to environmental and life danger. In this sense, accurate registration algorithms are essential in supporting Earth and medical scientists in their studies involving image aggregation and comparison. Research on diseases that remain largely mysterious, like Alzheimer for instance, could greatly benefit from the gains in terms of accuracy and computational time aimed in Dr. Enzo Ferrante’s project. Similarly, his innovative image registration methods and toolboxes will likely contribute to a better image-based understanding of Climate change and its consequences on the Earth’s surface.

When deep learning meets image registration

Explainable AI for healthcare: enabling a revolution

University of Oxford

Fairness in AI: ending bias and discrimination

University of Antwerp

Fulfilling the potential of AI: towards explainable deep learning

University of Antwerp