Abstracts

Rationale: Focal cortical dysplasia (FCD) is a highly epileptogenic developmental malformation (Barkovich AJ, Brain, 2012). On MRI, FCD type II is characterized by cortical thickening, blurring of the grey-white matter junction, and hyperintense signal. Although reliable detection of this lesion is critical for successful surgery, many lesions elude best-practice neuroimaging protocols (Bernasconi A, Nat Rev Neurol, 2011). The current work combined machine-learning techniques with surface-based MRI analysis in a series of consecutive patients with FCD type II, who were initially diagnosed as MRI-negative on routine radiological inspection. Methods: Based on 3T T1-weighted MRI (3D MPRAGE; voxel size: 1x1x1mm), we extracted cortical surfaces (Kim JS, Neuroimage, 2005) and measured cortical thickness, sulcal depth, and curvature as well as MRI intensity, and gradient features. We designed a two-step classification based on linear discriminant analysis. The first classifier was trained on patients' multivariate feature sets sampled from all lesional and randomly selected non-lesional surface points to generate surface probability maps for each individual to test. A subsequent classification pruned false positive findings based on cluster-wise Mahalanobis' distance of features. The classifier was trained and tested on 3.0T MRIs of 19 consecutive patients with MRI-negative FCD (15 were histologically confirmed), 24 healthy controls, and 11 disease controls (patients with temporal lobe epilepsy and histologically proven hippocampal sclerosis). Cross-validation followed a leave-one-out approach. Moreover, the ability to generalize our classifier to different data was assessed using a second dataset (14 patients with histologically confirmed FCD and 20 healthy controls) scanned at 1.5T. Results: The flow diagram (Fig 1) outlines the study design and results. Sensitivity was 74%, with 100% specificity (i.e., no lesions detected in healthy nor disease controls) on 3T dataset. In 50% of cases, a single cluster co-localized with the FCD lesion, while in the remaining a median of one extra-lesional cluster was found. An example of lesion detection is shown in Fig 2 Findings remained consistent when analysing only the subgroup of patients with histologically proven FCD. Using the 3T data as training set to classify the 1.5T dataset showed comparable performance (sensitivity: 71%, specificity: 95%). Conclusions: This study provides for the first time Class II diagnostic evidence that automated machine learning of MRI patterns accurately identifies FCD in patients with extra-temporal epilepsy initially diagnosed as MRI-negative. Our algorithm relies on surface-based multivariate pattern recognition that statistically combines morphology and intensity taking advantage of their covariance, thus unveiling sub-threshold tissue properties not readily identified on a single modality. The proposed method showed generalizability across cohorts, scanners, and field strengths. Machine learning may assist pre-surgical decision-making, by facilitating hypothesis formulation about the epileptogenic zone.