Robust and Generalisable Segmentation of Subtle Epilepsy-causing Lesions: a Graph Convolutional Approach

Hannah Spitzer, Mathilde Ripart, Abdulah Fawaz, Logan Z. J. Williams, Meld Project, Emma Robinson, Juan Eugenio Iglesias, Sophie Adler, Konrad Wagstyl

Abstract

Focal cortical dysplasia (FCD) is a leading cause of drug-resistant focal epilepsy, which can be cured by surgery. These lesions are extremely subtle and often missed even by expert neuroradiologists. "Ground truth" manual lesion masks are therefore expensive, limited and have large inter-rater variability. Existing FCD detection methods are limited by high numbers of false positive predictions, primarily due to vertex- or patch-based approaches that lack whole-brain context. Here, we propose to approach the problem as semantic segmentation using graph convolutional networks (GCN), which allows our model to learn spatial relationships between brain regions. To address the specific challenges of FCD identification, our proposed model includes an auxiliary loss to predict distance from the lesion to reduce false positives and a weak supervision classification loss to facilitate learning from uncertain lesion masks. On a multi-centre dataset of 1015 participants with surface-based features and manual lesion masks from structural MRI data, the proposed GCN achieved an AUC of 0.74, a significant improvement against a previously used vertex-wise multi-layer perceptron (MLP) classifier (AUC 0.64). With sensitivity thresholded at 67%, the GCN had a specificity of 71% in comparison to 49% when using the MLP. 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.

Introduction

Focal Cortical Dysplasia (FCD) is a primary driver of drug-resistant epilepsy. While surgical intervention offers a potential cure, identifying these lesions is notoriously difficult. Expert neuroradiologists fail to detect FCD in up to 43% of cases, categorizing these scans as "MRI-negative." Current automated detection tools, typically based on Multi-Layer Perceptrons (MLPs) or standard Convolutional Neural Networks (CNNs), suffer from a "local vision" problem. By analyzing vertices or small patches in isolation, these models lack the necessary whole-brain context, resulting in high false-positive rates that hinder clinical utility.

Methods

We propose a surface-based semantic segmentation approach utilizing a Graph Convolutional Network (GCN). The architecture adapts the state-of-the-art nnU-Net for spherical meshes (icospheres) using spiral convolutions. Unlike grid-based kernels, spiral convolutions define filters by following an outward path from a central node, allowing the model to operate directly on irregular cortical meshes.

Overview of the GCN architecture, including spiral convolutions and auxiliary loss components.

To improve robustness, the model incorporates two auxiliary losses:

  1. Distance Regression: The network predicts the geodesic distance of every point from the lesion boundary, penalizing false positives far from the target.
  2. Weak Supervision: A hemisphere-level classification loss helps the model learn from potentially noisy or uncertain manual masks.

The study leveraged the MELD (Multi-centre Epilepsy Lesion Detection) dataset, comprising 1015 participants (618 patients, 397 controls) from 22 international centers. Features included 11 surface-based metrics such as cortical thickness and FLAIR intensity, all registered to a symmetrical "fsaverage" template.

Results

The proposed GCN model (GC-nnU-Net+dc) demonstrated superior performance over the baseline MLP. The Area Under the Curve (AUC) increased from 0.64 to 0.74. Critically, at a fixed sensitivity of 67%, the GCN achieved a specificity of 71%, compared to just 49% for the MLP.

Comparison of lesion segmentation performance and false positive reduction between the MLP and the proposed GCN model.

Qualitative analysis shows that the GCN produces more contiguous and accurate lesion masks. Quantitatively, the model significantly reduced the number of false-positive clusters in both patients and healthy controls. This reduction is essential for clinical adoption, as it minimizes the number of areas requiring expert review and builds trust in AI-assisted diagnostics.

Discussion and Conclusion

The success of the GCN highlights that FCD detection relies heavily on global context—how a region appears relative to the entire brain. The inclusion of distance regression as a regularizer was pivotal in addressing the false-positive challenge. Furthermore, analysis suggested that performance plateaus around 220 subjects, indicating that data diversity and quality are more critical than sheer volume for future multi-center efforts.

While promising, the model remains limited by the inherent noise in human-annotated "ground truth" masks. Future work will focus on prospective validation in real-world radiological workflows and extending the graph-based approach to other epilepsy-causing pathologies.

Summary updated

Abstract

Focal cortical dysplasia (FCD) is a leading cause of drug-resistant focal epilepsy, which can be cured by surgery. These lesions are extremely subtle and often missed even by expert neuroradiologists. "Ground truth" manual lesion masks are therefore expensive, limited and have large inter-rater variability. Existing FCD detection methods are limited by high numbers of false positive predictions, primarily due to vertex- or patch-based approaches that lack whole-brain context. Here, we propose to approach the problem as semantic segmentation using graph convolutional networks (GCN), which allows our model to learn spatial relationships between brain regions. To address the specific challenges of FCD identification, our proposed model includes an auxiliary loss to predict distance from the lesion to reduce false positives and a weak supervision classification loss to facilitate learning from uncertain lesion masks. On a multi-centre dataset of 1015 participants with surface-based features and manual lesion masks from structural MRI data, the proposed GCN achieved an AUC of 0.74, a significant improvement against a previously used vertex-wise multi-layer perceptron (MLP) classifier (AUC 0.64). With sensitivity thresholded at 67%, the GCN had a specificity of 71% in comparison to 49% when using the MLP. 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.
BibTeX
@misc{spitzer2023robust,
  author = {Hannah Spitzer and Mathilde Ripart and Abdulah Fawaz and Logan Z. J. Williams and Meld Project and Emma Robinson and Juan Eugenio Iglesias and Sophie Adler and Konrad Wagstyl},
  title = {{Robust and Generalisable Segmentation of Subtle Epilepsy-causing Lesions: a Graph Convolutional Approach}},
  year = {2023},
  eprint = {2306.01375},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2306.01375},
}
Figure