Automated detection of lesions from MRI in children with focal epilepsy: a pilot study
Poster No:
420
Submission Type:
Abstract Submission
Authors:
Mathilde Ripart1, Maria Eriksson1,2, Rory Piper1,2, Chris Clark1, Felice D'Arco2, Kshitij Mankad2, Torsten Baldeweg1, Sophie Adler1, Konrad Wagstyl3
Institutions:
1UCL GOSH Institute of Child Health, London, United Kingdom, 2Great Ormond Street Hospital, London, United Kingdom, 3UCL Wellcome Centre for Human Neuroimaging, London, United Kingdom
First Author:
Co-Author(s):
Maria Eriksson
UCL GOSH Institute of Child Health|Great Ormond Street Hospital
London, United Kingdom|London, United Kingdom
UCL GOSH Institute of Child Health|Great Ormond Street Hospital
London, United Kingdom|London, United Kingdom
Rory Piper
UCL GOSH Institute of Child Health|Great Ormond Street Hospital
London, United Kingdom|London, United Kingdom
UCL GOSH Institute of Child Health|Great Ormond Street Hospital
London, United Kingdom|London, United Kingdom
Introduction:
Drug-resistant focal epilepsy can be caused by a broad range of structural brain abnormalities, from large tumours to subtle cortical malformations (Eriksson et al. 2023). It can be cured with resective epilepsy surgery provided that the abnormality is detected on MRI scans. AI models have recently been developed to aid the detection of specific epilepsy pathologies, such as focal cortical dysplasia (FCD) (Spitzer, Ripart et al. 2022). However, FCD represents less than half of the subtle epilepsy pathologies that can be missed on clinical review of MRI scans and there is an urgent clinical need to develop models capable of detecting a broader range of epilepsy pathologies. In this study, we investigate whether a single classifier can segment multiple pathological causes of focal epilepsy.
Methods:
This study was performed on 111 paediatric patients with a histologically confirmed cause of focal epilepsy and 90 paediatric healthy controls acquired for research purposes, from Great Ormond Street Hospital in London, UK. Patients' histopathologies included FCD, hippocampal sclerosis (HS), low-grade epilepsy-associated tumours (LEAT) – including dysembryoplastic neuroepithelial tumours (DNET), ganglioglioma – and other pathologies such as ganglioglioma, hypothalamic hamartoma, cavernoma non-diagnostic histopathologies. All participants had a 3D T1w scan acquired on a 1.5T or 3T MRI scanner, which was acquired preoperatively in patients and used to draw manual lesion masks by one of two expert radiologists. The cohort was split into training/validation (89 patients, 71 controls) and test datasets (22 patients, 19 controls) (Table 1). We used nnU-Net (Isensee et al. 2021), an open-source deep learning model specialised for robust performance in biomedical image segmentation, to segment focal epilepsy abnormalities from the T1w MRI scans using the manual lesion mask as ground truth. The nnU-Net was trained on the training/validation dataset for 1000 epochs with the '3d_lowres' configuration. It was evaluated on the test dataset for its sensitivity in detecting lesions (i.e. overlap between the prediction and the manual lesion mask) and specificity in controls (i.e. no prediction).
Results:
The model detected 14 out of the 22 focal epilepsy abnormalities in patients (64% sensitivity). It accurately detected 67% of FCD (n=6/9), 75% of HS (n=3/4), 80% of LEAT (n=4/5) and 25% of the other pathologies (n=1/4). Notably, the model accurately detected two out of the four abnormalities previously reported MRI-negative (one FCD and one LEAT). Patients had no more than one false-positive clustered prediction, and the model accurately predicted no putative lesions in 18 out of 19 healthy controls (95% specificity). Figure 1 depicts examples of three accurate predictions in patients with different pathologies: FCD 2B, HS and DNET. The radiological characteristics visibly differ between the three pathologies (Panel A). Nonetheless, the model was able to segment these three pathologies with a good overlap with the manual lesion masks (Panel B).
Conclusions:
We demonstrate that a single deep-learning model can segment multiple focal epilepsy pathologies on T1w MRI scans using a modest cohort of paediatric patients for training. This work paves the way for a larger, multi-centre study, aiming to develop a robust automated lesion segmentation tool that could help in the presurgical planning of patients with focal epilepsy.
Disorders of the Nervous System:
Neurodevelopmental/ Early Life (eg. ADHD, autism) 1
Modeling and Analysis Methods:
Segmentation and Parcellation 2
Novel Imaging Acquisition Methods:
Anatomical MRI
Keywords:
Epilepsy
Machine Learning
MRI
Open-Source Code
PEDIATRIC
Segmentation
STRUCTURAL MRI
1|2Indicates the priority used for review
Provide references using author date format
Isensee et al. 2021. “nnU-Net: A Self-Configuring Method for Deep Learning-Based Biomedical Image Segmentation.” Nature Methods 18 (2): 203–11.
Spitzer, Ripart et al. 2022. “Interpretable Surface-Based Detection of Focal Cortical Dysplasias: A Multi-Centre Epilepsy Lesion Detection Study.” Brain: A Journal of Neurology, August.