Abstracts

Rationale: Focal cortical dysplasia (FCD) is one of the most common pathologies for medically intractable focal epilepsy. Conventional weighted MRI can be limited in detecting or characterizing subtle FCD, partly due to the lack of quantitative measurements for tissue properties. Recently, a novel MR fingerprinting (MRF) technique has been developed (Ma et al, Nature 2013). MRF T₁ and T₂ tissue property maps were highly sensitive to epileptic lesions in prior studies (Ma et al, JMRI 2019; Liao et al, Radiology 2018). Here, we adopted a MRF-based radiomics feature extraction approach to characterize FCD lesions.

Methods: A 3D whole-brain MRF scan (Ma et al, JMRI 2019) was acquired from 43 healthy controls (HCs) and 21 epilepsy patients with pathologically confirmed FCD (12 type II and 9 type I) at 3T (1.2 mm³ isotropic voxels, scan time 10.4 min). MRF T₁ map, T₂ map and synthetic T_1w images were reconstructed based on a predefined dictionary. 3D Lesion ROIs were manually created from T_1w images. The T_1w images of patients and HCs were normalized to the MNI space. The warping information from T_1w images was applied to the T₁, T₂ maps and ROIs. CSF was segmented and removed from ROIs. Ninety-six radiomics features in the ROIs were extracted from T₁ and T₂ maps using the radiomics toolbox in MATLAB. The same ROIs were applied to HCs to extract average values of features at the matched locations. Features were extracted from each slice of 3D ROIs for data augmentation. Two logistic regression ML models were generated: FCD vs HC and FCD type II vs FCD type I. Datasets were divided into a training set (80%) and a testing set (20%). In the training set, we performed backward elimination for the optimal feature selection based on p-values between groups. Features with the highest average accuracy from 10-fold cross-validation were selected as the optimal features. We then validated the models with these optimal features in the testing set. Performance metrics were averaged from 3 shuffling datasets to minimize bias.

Results: Figure 1 shows representative MRF images for FCD IIB, IIA and I. To classify FCD from HC,s the MRF radiomics ML model with the optimal 30 features showed high AUC, accuracy, sensitivity and specificity of 0.86±0.02, 84±1%, 87±1% and 80±2%, respectively, in the training set. In the testing set, the model showed stable performance of 0.86±0.02, 81±2%, 86±3%, 76±4% for the respective values. For classifying type II from type I, Figure 2A shows the results of backward elimination in the training set, for which the model with the optimal 47 features showed AUC, accuracy, sensitivity and specificity of 0.80±0.00, 81±0%, 78±3%, and 84±3%, respectively. In the testing set, the model also showed stable performance of 0.88±0.02, 82±3%, 84±8%, and 81±3% for the respective values (ROC in Figure 2B).

Conclusions: Our data showed high accuracies of the novel MRF-based radiomics ML models to classify FCD from healthy tissue, and to classify FCD type II from type I. This work demonstrates the potential of using MRF-based radiomics to aid the noninvasive presurgical evaluation of epileptic individuals.

Funding: Please list any funding that was received in support of this abstract.: NIH R01 NS109439.