Abstracts

RATIONALE: High-resolution MRI of the brain has made it possible to identify focal cortical dysplasia (FCD) in an increasing number of patients. However, in many patients, lesions of FCD are characterized by minor structural abnormalities that go unrecognized by standard radiological analysis. We previously developed a voxel-based image processing method including first-order texture and morphological analysis using features that were chosen to model in vivo the pathological characteristics of FCD (A. Bernasconi et al., Annals of Neurology, in press). The purpose of the current study was to implement an objective method of detection of FCD lesions by developing a fully automated classifier based on the same image processing techniques.
METHODS: We selected 15 patients with histologically proven FCD. Image processing features were calculated for each voxel in the 3D T1-weighted MRI, generating three feature maps: (i) gray matter thickness map to model cortical thickening; (ii) gradient map to model blurring of the GM-WM junction; (iii) relative intensity map to model hyperintense signal within the lesion. We developed a Bayesian classifier to identify four categories: lesion, gray matter, white matter and cerebrospinal fluid. We selected five patients with FCD to use in the construction of a training set. T1-MRI and the three feature maps of these patients were used to train the classifier. A separate testing set consisted of the MRI and feature maps of the remaining 10 patients and a set of 10 controls. The output of the classifier was a 3D volume for each patient, segmented into the four categories listed above (Figure shows a representative slice). Lesions were independently visually identified and manually segmented on T1-weighted 3D MRI prior to the automatic classification for later use in validation.
RESULTS: In all ten patients the lesion was classified correctly, indicated by agreement with the location of the manually segmented lesion on the T1-MRI. All subjects had voxels that were misclassified as lesion. However, given their scattered distribution over the entire brain, these were easily distinguishable from FCD.
CONCLUSIONS: In this study, we demonstrate feasibility and proof of concept of a novel and fully automated method for detection of FCD on MRI. Our method demonstrates that additional information of use in identifying FCD can be extracted from a T1-weighted MRI. This technique could further improve the ability to detect dysplastic lesions in patients with intractable partial epilepsy.[figure]