Surface-based segmentation of Focal Cortical Dysplasias using Graph Neural Networks: a MELD study

2009 

Abstract Submission 

Konrad Wagstyl¹, Emma Robinson², Sophie Adler¹, Mathilde Ripart³, Logan Williams², Juan Eugenio Iglesias⁴, Hannah Spitzer⁵, MELD Project¹, Abdulah Fawaz⁶

¹UCL, London, London, ²King's College London, London, London, ³UCL, London, Select a State, ⁴Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, MA, ⁵Institute of Stroke and Dementia Research, Munich, Bavaria, ⁶KCL, London, London

Konrad Wagstyl, Dr.  
UCL
London, London

Emma Robinson, Dr  
King's College London
London, London

Sophie Adler, Dr.  
UCL
London, London

Mathilde Ripart  
UCL
London, Select a State

Logan Williams  
King's College London
London, London

Juan Eugenio Iglesias, Ph.D.  
Athinoula A. Martinos Center for Biomedical Imaging
Charlestown, MA

Hannah Spitzer  
Institute of Stroke and Dementia Research
Munich, Bavaria

MELD Project  
UCL
London, London

Abdulah Fawaz  
KCL
London, London

Focal cortical dysplasia (FCD) is a common cause of drug-resistant epilepsy, and accurate detection on MRI is critical for presurgical planning(Téllez-Zenteno et al., 2010). However, MRI identification of these subtle lesions remains a challenge. Previous FCD detection methods have been prone to high numbers of false positives due to their inability to take the entire cortex into account(David et al., 2021; Gill et al., 2021; Spitzer et al., 2022). This Multicentre Epilepsy Lesion Detection (MELD) project study aimed to develop a whole-brain graph neural network (GNN) for segmenting FCDs.

The MELD cohort includes 618 patients with FCD and 397 controls, used to train and test a novel Graph UNet model architecture (https://github.com/MELDProject/meld_graph). The cortical mesh was represented as a graph, treating vertices as nodes connected to neighbouring vertices by edges, enabling the network to learn spatial relationships through spiral convolutional filters. We trained the network on 278 patients and 180 controls to segment lesions with four parallel tasks: lesion segmentation, prediction of geodesic distance from the lesion, object detection and classification of lesional examples. These last two tasks were designed to mitigate uncertainty in manually delineated lesion masks. The network was evaluated on a withheld test dataset (260 patients, 193 controls) for its sensitivity in detecting lesions in patients (i.e. overlap between prediction and ground truth) and specificity in controls (i.e. no false positives).

On the withheld test cohort, the MELD Graph model achieved a sensitivity of 67% in patients, with a specificity of 76% in controls, a significant gain in specificity in controls against patch-based approaches on the same dataset (sensitivity 67%, specificity 49%). The MELD Graph model increased the positive predictive value of a prediction being a lesion from 0.4 to 0.7. Interpretable lesion reports characterise lesion location, salient feature abnormalities, and overall prediction confidence.

Our study demonstrates the utility of GNNs for FCD segmentation in MRI scans. The fully-trained GNN substantially improved on previous patch-based approaches. This improvement in specificity is vital for clinical integration of lesion-detection tools into the radiological workflow, through increasing clinical confidence in the use of AI radiological adjuncts and reducing the number of areas requiring expert review.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ²

Segmentation and Parcellation ¹

Cortex

Epilepsy

Neurological

Open Data

Open-Source Code

Open-Source Software

Pediatric Disorders

Segmentation