Using artificial intelligence to mitigate scanner bias in brain disorders

Brain-based disorders represent 14% of the global burden of disease. Artificial Intelligence (AI) applications to neuroimaging have the potential to improve the detection and treatment of these disorders. This potential has been demonstrated in controlled research setups, but the real-world clinical implementation of these techniques is limited by differences in scanning equipment and protocols. These differences, known as scanner bias, means that diagnostic and prognostic AI applications developed using images from a certain machine tend to perform poorly when applied to images from different machines. There is preliminary evidence that some AI-based approaches can mitigate this problem, however, they are limited to healthy subjects. We will adapt and validate Neuroharmony, an AI-based solution for this problem, using data from clinical populations with psychiatric or neurological disorders obtained across multiple machines. The results obtained here will improve Neuroharmony, validate it for clinical application and make it available to the research and clinical communities. This tool can help bridge the gap between AI-based research and clinical practice.

In 2017 brain-based disorders represented 14% of the global burden of disease. Imaging-based analyses have the potential to provide reliable diagnostic and prognostic markers for brain-related disorders. A number of successful applications to brain disorders have been reported. However, a critical challenge in bringing these machine learning applications to clinical practice is that the use of different scanners and acquisition protocols results in inconsistent measures of the brain. In some cases, scanner-related variability can blur or even eliminate clinically relevant information from the scans. A number of methods for mitigating inter-scanner bias, commonly known as data harmonisation, have been proposed. However, the consolidated harmonisation methods have limited translational value as they do not allow the harmonisation of individual images coming from a new unseen scanner, which is essential for a real-world application of machine learning. They typically need a statistically significant sample as a reference, therefore they cannot be implemented in a cross-validation design. To overcome this challenge, we developed Neuroharmony, a tool that can harmonise single images from unseen/unknown scanners based on a set of image quality metrics (e.g. intrinsic characteristics like signal-to-noise ratio, contrast-to-noise ratio, etc.) which can be extracted from individual images without requiring a statistically representative sample from each new scanner of interest. Neuroharmony was proven effective in reducing systematic scanner-related bias from the distribution of brain volumes of healthy controls. However, its generalizability to patient data is yet to be examined. In this project, we aim to demonstrate that Neuroharmony can be used to harmonise data collected from clinical samples while preserving disease-related variability.