Abstracts

Rationale: Focal cortical dysplasia (FCD) is a congenital abnormality of cortical development and a leading cause of surgically remediable drug resistant epilepsy. Machine learning offers a powerful framework to develop automated and individualized tools that aid the detection of lesions (Adler, Wagstyl et al., NICL, 2017; Jin et al., Epilepsia, 2018) . However, machine learning continues to improve with increasing numbers of examples necessitating multi-centre collaboration (Figure 1B). Here we have created an open infrastructure, with freely accessible pipelines and protocols for parallel post-processing of clinical MRI scans across multiple sites internationally. This approach equips individual sites to replicate our pipeline locally and standardizes preprocessing, enabling multi-centre data sharing. Methods: The complete MELD processing pipeline was developed into five protocols. These were uploaded to the open-science platform, Protocols.io (https://www.protocols.io/researchers/meld-project/protocols). These are step-by-step instructions for each participating site to follow. All code required is available at https://github.com/MELDProject/meld.Protocol 1 includes the inclusion and exclusion criteria for the study alongside how to fill in the demographic information.Protocol 2 contains instructions for FreeSurfer reconstructions of T1 and FLAIR MRI data.Protocol 3 details how to conduct quality control of the FreeSurfer reconstructions.Protocol 4 describes how to create masks of the FCDs, which is necessary for supervised learning by the neural network classifier.Protocol 5 is the post-processing pipeline. This protocol explains how to run the code that samples numerous surface-based features including cortical thickness, grey-white matter intensity contrast, curvature, sulcal depth, intrinsic curvature and FLAIR signal. These features are registered to a symmetrical template (fsaverage_sym). Intra- and inter-subject normalisation is conducted on each feature. Finally, per-vertex features and the lesion masks are stored for all participants in a large anonymised data matrix.The final stage of the pipeline is to share the anonymised data matrices. This will enable us to improve the performance of our existing neural network classifier by training on anonymised clinical data from 100s of patients and controls world-wide (Figure 1A), crucially without necessitating the sharing of clinical MRI data. The distributed nature of the processing enables each centre to train and evaluate a classifier locally. Results: The protocols were developed and tested on MRI data from paediatric patients at Great Ormond Street Hospital, UK. The protocols have now been piloted at an independent collaborating centre, the Florey Institute of Neuroscience and Mental Health, Australia. Troubleshooting, involving clarifying instructions, fixing environment and path errors, and correcting coding bugs, was facilitated through providing comments on Protocols.io. This open-science framework has enabled version 2 of the protocols to be released, which can now be used by all of the fifteen collaborating centres as well as any other interested group. Conclusions: A consortium of epilepsy centres (Figure 1A; https://meldproject.github.io/groups/) have joined a group endeavour to create open-access, robust and generalisable tools for FCD detection. Protocols.io provides an interactive, open-science platform to develop and test protocols and code for FCD detection. MELD offers a new paradigm for conducting clinical data-driven research across multiple sites and training clinicians and scientists worldwide to conduct surface-based MRI research. Funding: We thank the Rosetrees Trust for their generous support.