Developing environmentally sustainable best practices for human brain imaging

Human brain imaging techniques such as Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET) and Computerised Tomography (CT) are invaluable healthcare research tools. However, increasingly, their acquisition, storage, and analysis are generating substantial environmental costs.

At present, the brain imaging research community has little awareness of these costs, and little to no understanding of how to reduce them. Taking energy demands as an example, we do not know exactly how much energy is used over a dataset lifecycle. This means that it is difficult to identify what precise steps in the brain imaging research process to target for energy reduction. There are, however, emerging tools we can draw upon from parallel fields that are more advanced in calculating carbon footprints of research procedures - in particular, from the field of machine learning (ML).

This project will focus on assessing the carbon impacts associated with brain imaging research, using existing approaches to carbon tracking ML models to create a neuroimaging carbon tracker tool. This tool will be embedded within commonly used neuroimaging software and research tools, enabling the field to measure the carbon footprint of brain imaging research procedures.

We will then use the neuroimaging carbon tracker to measure the carbon footprint of a variety of brain imaging research procedures and examples, including using publicly available software and tools to enable wider adoption. Having identified the precise procedural steps that drive neuroimaging research footprints, we will then develop optimisation strategies to reduce energy consumption, and thereby, carbon emissions. For example, we will investigate how to remove unnecessary analysis steps, and how to store data optimally. Ultimately, this will enable us to generate 'best practice' guidance for researchers in neuroscience, to enable them to adopt more sustainable research procedures. We will share this guidance with the research community via open source software tools, an open access '10 simple steps' journal publication, conferences, and a cross-sectoral workshop with colleagues across healthcare sciences.

By raising awareness of the carbon costs of human brain imaging, and providing specific recommendations on how to reduce this, we will facilitate human brain imagers, and colleagues in related fields, to minimise the environmental footprint of imaging-based life sciences research.

The rapid increase in the accessibility to high performance computing has accelerated large-scale analysis of complex medical imaging data. This has particularly impacted the study of the human brain by enabling application of novel computational methods to imaging modalities such as MRI, PET and CT. The computational study of the human brain has, in turn, led to the development of elaborate pipelines operating on vast amounts of data which are being adopted globally. As a consequence, the computational costs of such neuroimaging pipelines have increased considerably. The increasing storage and computational requirements of neuroimaging pipelines translate into significant increase in the corresponding energy- and carbon- costs.

In this project, we aim to assess the carbon footprint of computational neuroimaging pipelines and present basic measures to reduce these costs in some instances. As the first objective, we will adapt existing carbon-tracking tools for neuroimaging pipelines, which we hope will first provide an insight into the carbon-intensity of these pipelines. We will build tools to measure the energy costs of running specialised computation infrastructure used in research data centres. We also aim to 1.) develop a slurm-based job scheduler that can utilise the instantaneous carbon intensities of local power grids to modulate and reduce the energy costs of neuroimaging pipelines 2.) investigate the usefulness of image derived phenotypes in tasks where using massive amounts of raw image data can be computationally expensive. 3.) systematically explore the loss in precision compared to the gains in carbon efficiency. Finally, as neuroimaging pipelines comprise several intermediate steps, we also aim to assess different alternatives and configurations of these steps. This we hope will serve as a guiding framework when deciding the trade-off between a specific choice of method and its corresponding carbon cost.