Abstracts

Rationale: Focal cortical dysplasia (FCD) is a congenital abnormality of cortical development and a leading cause of surgically remediable drug resistant epilepsy. MRI has played a major role in the evaluation of patients; yet, significant proportions of lesions remain undetected by conventional image analysis. Machine learning offers a powerful framework to develop automated and individualised clinical tools that may improve the detection of lesions and prediction of clinically relevant outcome. To date, surface-based automated lesion detection studies have been single-centre and it is unclear how generalisable the developed frameworks and tools are. Furthermore, machine learning continues to improve with increasing numbers of examples. Here, we create an international collaboration for lesion detection and develop normalisation techniques for the incorporation of data and sequences from multiple sites. Methods: T1 and FLAIR sequences in FCD patients and matched controls are acquired at each site (Great Ormond Street Hospital for Children, UK; National Hospital for Neurology and Neurosurgery, UK; Cleveland Clinic, USA; Second Affiliated Hospital of Zhejiang University, China; Beijing Tiantan Hospital of Capital Medical University, China). Cortical reconstructions are generated using FreeSurfer software. The following surface-based metrics are calculated per vertex across the cortical surface: cortical thickness, grey-white matter blurring, FLAIR signal intensity, sulcal depth, curvature. To compare patients across different ages, scanners and sequence parameters each feature undergoes intra- and inter-subject normalisation by site-specific controls. Intra-subject normalisation accounts for age-specific differences in features, whereas inter-subject normalisation for regional variations in a feature across the cortical surface. Features can then be incorporated into a neural network that can be trained to identify lesional vertices. All scripts are freely available at https://github.com/kwagstyl/FCDdetection/. Results: Before normalisation, subjects of different ages who have also been scanned on different scanners have large inter-individual differences in the distributions of each surface-based feature (Figure 1A). For example, younger patients have thicker cortices. After intra- and intra subject normalisation – all features have a normal distribution centred around zero with a standard deviation of one (Figure 1B). Conclusions: A consortium of epilepsy centres have joined a group endeavour to create open-access, robust and generalisable tools for FCD detection. Two-stage, intra- and inter- subject, normalisation of surface-based feature maps ensures that all patient and control data regardless of age, scanner or sequence parameters are comparable. This normalisation is integral for multi-centre structural MRI studies. Feature maps can now be incorporated into machine learning algorithms for FCD lesion detection. Our goal is to make these tools generalisable to any site, we encourage interested groups to contact us to join the MELD project. Funding: Epilepsy SocietyGreat Ormond Street Hospital for Children NHS Foundation Trust Epilepsy Reseach UKGreat Ormond Street Hospital Children's CharitySA was funded by the Rosetrees Trust (A711)DC was funded by Action Medical Research (GN2214) GPW was supported by an MRC Clinician Scientist Fellowship (MR/M00841X/1)