Introduction: Focal cortical dysplasia (FCD) is a significant cause of drug-resistant epilepsy, often eluding detection on conventional MRI. Accurate automated detection is crucial, as surgical resection of FCD can lead to seizure freedom in many patients. This study presents a novel, multi-step approach for identifying FCD lesions on MRI. Rather than relying on a single model, this method combines automated segmentation, radiomics feature extraction, and a space-filling algorithm that integrates every voxel indicating structural irregularity into a coherent lesion. The approach was tested on an OpenNeuro dataset with MRI data from 85 patients with FCD type II and 85 healthy controls.
Methods: The analysis began with automated segmentation on T1-weighted and FLAIR images, delineating cortical and subcortical structures. Radiomic features—including textural, geometric, and intensity-based metrics—were extracted from these segmented regions to capture attributes characteristic of FCD. Finally, a space-filling algorithm was applied to aggregate each voxel indicating irregularity into a unified lesion. Classification performance was assessed using a random forest classifier with cross-validation.
Results: The proposed method achieved an overall accuracy of 93%, with a sensitivity of 95% and specificity of 90% in identifying FCD lesions. Radiomic features provided strong differentiation: kurtosis values averaged 1.32 in FCD regions versus 0.85 in controls (p < 0.01), while entropy and fractal dimension were also higher in FCD lesions, indicating increased structural irregularity (p < 0.05). The space-filling algorithm notably improved sensitivity for lesions under 1 cm³, enhancing lesion coherence by aggregating all voxels of irregularity. This method successfully localized lesions in 81 out of 85 patients.
Conclusion : This multi-step detection method, integrating segmentation, radiomics, and a space-filling algorithm, demonstrates high accuracy and sensitivity in detecting FCD on MRI, with strong explainability through interpretable radiomic features and effective utility in identifying small lesions. The approach offers a reliable, reproducible tool for pre-surgical evaluation and clinical decision-making in epilepsy management. Further validation on larger, diverse datasets is recommended to establish its clinical utility.
.jpg)
