Abstracts

Rationale: Drug-resistant focal epilepsy (DRFE) is commonly caused by structural brain abnormalities that can be amenable to epilepsy surgery. However, these lesions can be subtle and overlooked on radiological review. Individual AI models have been developed to detect specific epilepsy pathologies, such as focal cortical dysplasia or hippocampal sclerosis [1,2]. However, at the time of MRI review, the underlying pathology is not known. There is an urgent clinical need to develop models capable of detecting a broad range of focal epilepsy pathologies. In this study, we leverage the MELD Focal Epilepsy dataset, the largest collection of 3D MRI scans in patients with DRFE, to investigate whether a single classifier can segment a variety of structural causes of focal epilepsy.


Methods: The study included 1181 patients with DRFE and 1009 controls, from 18 epilepsy centres worldwide. Pathologies included focal cortical dysplasia (FCD), hippocampal sclerosis (HS), low-grade epilepsy-associated tumours (LEAT) (including DNET and gangliogliomas), hypothalamic hamartoma (HH), cavernoma (CAV), polymicrogyria (PMG), periventricular nodular heterotopia (PNH) and other pathologies, such as MOGHE. All participants had a 3D T1w scan acquired at 1.5T or 3T. The cohort was split into 80% train-val and 20% test datasets (Table 1). For patients with hippocampal lesions, automated hippocampal segmentation from Synthseg [3] was used as the lesion mask. For all other lesions, experts from each center manually drew the ground truth lesion mask on the preoperative T1w scan. T1w MRI scans were used to train a nnU-Net model [4] to segment focal lesions. The model was evaluated on the test dataset for its sensitivity in detecting lesions (i.e. minimum 1 voxel overlap between the prediction and the lesion mask) and specificity in controls (i.e. no prediction).


Results: The model detected 170 out of the 232 focal epilepsy abnormalities (73% sensitivity) in patients (Table 1). It accurately detected 63% of FCD, 91% of HS, 80% of LEAT, 60% of CAV, 67% of PMG and 50% of HH, PNH and other pathologies. Notably, the model accurately detected 45% of lesions previously reported “MRI-negative”. Model specificity was 90%. Figure 1 depicts examples of seven accurate predictions in patients with different underlying pathologies.


Conclusions: We demonstrate that a single deep-learning model can segment a variety of pathologies associated with focal epilepsy on T1w MRI, on an heterogeneous cohort, representing a range of ages, countries and MRI scanners. The algorithm successfully detected 45% of MRI-negative lesions, evidence of its potential utility as a radiological adjunct. This work represents a significant step forward in the development of a robust automated lesion segmentation tool that could help in the presurgical planning for patients with focal epilepsy. Future work, will involve increasing the sample size of the rarer pathologies and incorporating multimodal FLAIR and T2 MRI data.


Funding: The MELD Project, M.R. and S.A. are supported by the Rosetrees Trust (A2665) and Epilepsy Research Institute (P2208). K.W. is supported by the Wellcome Trust.