Precision Modelling of Cortical Variation and its Association with Neurological/Psychiatric disease

The aim of this proposal is to develop a novel medical imaging support tool to significantly improve rates of detection, of types of subtle brain abnormality, which give rise to complex brain conditions. Specifically, we are seeking to develop tools that improve the accuracy with which we can compare brain scans across populations. This will make it much easier to tell the difference between healthy and atypical brains, or detect diseased tissue.

The reason that this is challenging is because brains are extremely complex, made of billions of cells, and each one can look very different. This makes it hard to build a single model of what "healthy" brains should look like, and as a result it becomes very difficult to spot evidence of disease.

These challenges mean that radiologists require years of experience, reviewing countless examples, before they can reliably spot subtle brain abnormalities, and even so, for diseases such as focal childhood epilepsies, up to 30% of cases evade detection. For similar reasons, automated tools often also struggle: appearance of scans varies so extensively that simplifying assumptions must be made leading to coarse solutions.

The largest assumption is that all brains share a common organisational blueprint, where areas of the brain responsible for different functions appear in the same order. Such that if each brain scan was a jigsaw, with each piece a region, the shapes might change but they would in go together in the same way. However, in reality brains vary topographically, which means that areas representing different functions (such as language) can swap location. Methods assuming otherwise end up comparing completely different areas of the brain across individuals. Each area may look very different, with different definitions of what is normal. As a result, this leads to confusion, limiting the ability of any method to detect signs of disease.

In the past, methods were particularly limited as they built their model of regional organisation based simply on patterns of brain folding. However, it turns out that shape is a fairly coarse and non-specific model of brain organisation, and that brains often have very different patterns of brain folding for the same functional region.

Recently we developed a novel open-access tool, which instead learns how to map brains onto a model which takes into account, not just shape but also function, and other aspects of brain organisation (Robinson Neuroimage 2014, 2018). This has led to new, more accurate, models of cortical organisation (Glasser Nature 2016) and development (Garcia PNAS 2018, O'Muircheartaigh Brain 2020) and improved understanding of the links between brain organisation and behaviour (Bijsterbosch Elife 2018).
Now we propose to extend this tool, to account for variation of brain shape and appearance in a way that reflects the natural variation seen from one individual to another. Rather than learn a single model of brain organisation we will learn a family of models (modes) that try to describe how our brains vary. These will capture all biologically relevant modes of variation, allowing individual brain scans to be compared, for a given location, only against others with a common organisational blueprint. In this way we will support much more detailed comparison, than was ever possible before.

We will validate the power of the approach through three studies: 1) finding the source of epileptic seizures in the brain (to support surgical planning); 2) predicting cognitive outcomes for babies with developmental brain conditions; 3) identifying biological markers in the brain that may help predict mental health conditions. Ultimately, these tools will support researchers, medical doctors and healthcare workers to build more sensitive predictive models, fine tuned to detect signs of abnormality within individual brains. This will improve screening detection rates and lead to more accurate diagnosis of all brain conditions.

The purpose of this grant is to develop urgently needed new technologies for image processing of brain scans to support precise comparisons of populations of brain scans and sensitive localisation of pathologies within individual scans.

It is motivated by the fact that current methods for brain image analysis largely ignore considerable variation in brain shape and cortical organisation across individuals in order to deliver well-posed optimisations, which allow reliable but coarse scale comparisons of data across populations. Specifically, these compare brains through use of image registration algorithms which support smooth and invertible spatial mappings of all data into a global average space, where images may be directly compared. Alternative approaches, such as deep learning (which avoids this) require enormous data-sets, and return limited explainability (leaving ethical concerns over bias and generalisation).

The key problem with image registration to a single, global template is that this typically ignores subtle sources of individual variation leading to (in some circumstances very) low powered studies of complex phenomena. This is because there is nothing inherent in these algorithms constraining atypical variants of brain architecture to be in correspondence.

In this proposal, we therefore seek to develop new techniques for brain image registration which seek to learn all clinically or behaviourally relevant patterns of topographic variation, with the intent of matching these during alignment. In this way, images will be brought into much closer correspondence than possible before, allowing precision generative modelling of population variation for detection of cortical abnormalities in small populations. These frameworks will be used to test and validate development of novel explainable deep-learning frameworks.