Abstracts

Rationale: Focal cortical dysplasias (FCDs) are an important cause of drug-resistant epilepsy that is amenable to surgery but are notoriously difficult to visually identify on structural MRI. Approaches to improving the detection of FCDs have involved improved scanner protocols and field strengths as well as automated volumetric and surface-based post-processing methods. However, the ability of automated methods to generalise and be translatable into presurgical evaluation has been harboured by small sample sizes and / or MRI scanners, limited inclusion criteria as well as code availability. We developed an open-source, surface-based machine-learning algorithm to automatically identify FCDs on heterogeneous structural MRI data from epilepsy surgery centres worldwide.

Methods: We collated a retrospective cohort of patients with epilepsy due to FCD and controls from 21 epilepsy centres. Multiple surface-based features, containing morphological and image intensity information, were extracted from T1w and FLAIR images. Features were smoothed, harmonized across sites using ComBat (Fortin et al., 2018), and normalized for intersubject and interregional morphological differences. Manual lesion masks were mapped to FreeSurfer surfaces. To account for the inherent uncertainty around lesion boundaries, border zones were created around each lesion mask extending approximately 10mm across the cortical surface. The full cohort of patients and controls were randomly assigned to either the train/val cohort or the test cohort. A multi-layer perceptron, with 2 hidden layers (40, 10) and 1 output node, was trained with a focal loss, to differentiate healthy and lesional vertices. The classifiers trained on each of the 10 folds in the train/val cohort were combined into an ensemble model, and evaluated on the test cohort. The main outcome measures were the sensitivity and specificity of the developed MELD algorithm to detect the FCDs after including a border zone around the lesion masks.

Results: 555 FCD patients and 390 controls were included. Examples of classifier outputs from patients with detected and not detected lesions is displayed in Figure 1. After including the border zone, the developed MELD surface-based algorithm had a sensitivity (prediction overlapping the lesion mask) of 68.0% and specificity (any false positive clusters on controls) of 33.3% on the withheld test cohort and a sensitivity of 72.3% and a specificity of 29.0% on the train/val cohort, retrained on the test cohort. In the gold-standard cohort of histopathologically confirmed FCD type II patients who were seizure free post-operatively, the sensitivity was 79.2% (Figure 2).

Conclusions: Through open infrastructure, freely accessible pipelines and protocols for parallel post-processing of clinical MRI scans and collaborative working practices we have developed a tool for the automated detection of FCDs that is robust to heterogeneous input data. This tool could assist the presurgical evaluation of patients with epilepsy.

Funding: Please list any funding that was received in support of this abstract.: Rosetrees Trust and Wellcome Trust.