Abstracts

Rationale: Standard MRI practice is often suboptimal for localizing the possible seizure onset zone (SOZ) across different histopathologies, possibly due to the cortical structures heterogeneously altered by the underlying epileptogenic lesions ranging from focal cortical dysplasia (FCD), tumor, and normal histology other than showing non-specific gliosis alone. The present study of children with focal epilepsy aimed at overcoming this limitation by using a state-of-the-art deep learning architecture, which disentangles various input data (multi-modal MRI and patient variables) and automatically determines what data are the most discriminative to improve SOZ localization. Our working hypothesis was that the abstraction of heterogeneous input data using the deep neural network (DNN) would accurately discriminate SOZ from non-SOZ regions defined by ictal ECoG recording. Methods: We studied 60 children with drug-resistant epilepsy (age: 5-18 years, 30 boys, 25 left temporal, 10 right temporal, 15 left extra-temporal, 10 right extra-temporal; 10 FCD type I, 18 FCD type II, 12 tumor, 20 non-specific/normal histology) in whom ECoG localized SOZ and resective surgery was subsequently performed. The total number of SOZ, interictal spiking, and non-epileptic electrode sites were 958, 1421, and 4069. Fourty multi-modal MRI features (from T1, T2, FLAIR and DWI sequences) and seven epilepsy variables (e.g., scalp EEG, age, gender, seizure frequency, seizure type, epilepsy duration, antiepileptic drugs) served as the input vector, xi, in our DNN (Figure 1). It classified xi into one of 5 classes, 'C1: SOZ with FCD I', 'C2: SOZ with FCD II', 'C3: SOZ with tumor', 'C4: SOZ with non-specific/normal histology', 'C5: non-SOZ with no structural lesion including spiking and non-epileptic electrodes'. Cross validation was employed to train and test all hidden layers to minimize focal loss in an end-to-end fashion. Finally, prediction probability values of a given xi beloinging to Cl=1-5 were evalued at the softmax layer to estimate seizure likelihood, mi (i.e., maximal prediction probability among four SOZ classes, C1-4).  Results: As shown in Figure 2, the DNN showed three hidden layers (J=3, M=100) which deeply trained non-linear relationships between input: xi and output: Cl=1-5 available in our study samples (n= 4,513, 70% of total 6,448 electrodes). The DNN achieved an accuracy of 92.1% in classifying individual xi as its ground truth class, Cl, yielding a high effect size, f=0.442, between SOZ and non-SOZ electrode sites. For comparison, the same dataset was analyzed by a shallow neural network (SNN) that is a subtype of DNN with a single hidden layer (J=1, M=100). The SNN provided an accuracy of 75.1% for SOZ localization in the test set (n=1,945, 30% of total 6,448 electrodes). Conclusions: Our empirical data suggest that DNN can be feasibly translated in the presurgical evaluation of children with drug-resistant focal epilepsy associated with diverse underlying etiologies. While taking into account the relevant patient variables, the DNN effectively learned the profiles of multi-modal MRI abnormalities highly specific to SOZ in our cohort. Funding: This study was supported by the following grants, R01 NS089659 (to J.J) and R01 NS064033 (to E.A) from the National Institute of Health (NIH).