Synopsis

Keywords: Neuro, Epilepsy, Ensemble deep learning, Localization of epileptogenic zone

We present a novel ensemble deep learning neural network of multi-scale deep MRI features that can non-invasively localize the epileptogenic zone (EZ) partially overlapping seizure onset zone (SOZ) and irritative zone (IZ) in children with medically intractable epilepsy. The presented network provided high balanced accuracy of 90% to predict SOZ and IZ in 3-fold cross-validation. It also yielded a new MRI marker of epileptogenicity providing a huge effect size between the ground-truth EZ and non-EZ (i.e., Cohen’s d = 2.73), suggesting its high translational value for accurately guiding invasive EEG practice to determine the boundaries of the EZ sites

Introduction

Surgical resection of the epileptogenic zone (EZ)¹ is an outstanding option for children whose seizures do not respond to medical therapy. Seizure onset zone (SOZ)² and irritative zone (IZ, defined as the region generating interictal spikes)³ on invasive EEG (iEEG) are electrophysiological biomarkers used to prospectively localize the EZ in clinical practice. Only 50-70% of patients achieve seizure freedom following the resection based on semiology, scalp EEG, iEEG, MRI, and other available neuroimaging workups, often due to incorrect or incomplete localization of the EZ⁴. This study proposes a novel ensemble learning neural network^5,6 combining a series of multi-scale residual networks (msResNet)⁷ to accurately localize the EZ sites by using their multi-scale deep features of clinical multi-modal MRI.

Methods

We utilized the 3T multi-modal MRI clinically acquired from our retrospective archive that were originally utilized for our previous study⁷ (n=41 children with intractable epilepsy who underwent two-stage resective surgery and achieved long-term seizure freedom after surgery, age: 9.9±5.6 years old, 22 boys). A model cohort (n=24 children) was used using a 3-fold cross-validation to train and test the proposed ensemble learning neural network for accurate prediction of C₁: SOZ, C₂: IZ, and C₃: Non-EZ in the end-to-end fashion (Fig. 1). A validation cohort (n=17 children) was then used to evaluate the reproducibility of the proposed network in an independent cohort. Briefly, the node-wise prediction was performed in the connectivity network, G = (Ω_i=1-500, A_{i=1-500,j=1-500}), where Ω_i defines the ith network node of the epileptogenic hemisphere in the Lausanne 2008 cortical parcellation atlas⁸ and A_ij represents pair-wise white matter pathway edge connecting Ω_i and Ω_j (e.g., connectivity strength, the total count of connecting tracts normalized by the total volume of Ω_i and Ω_j). At each node Ω_i, a surface laminar analysis^7,9 was applied to extract a multi-modal MRI feature vector X_i sized by 1 x 1,800 consisting of I: relative intensity (RI) values of T1-weighted, T2-weighted, FLAIR, apparent diffusion coefficient (ADC), fractional anisotropy (FA) at two gray matter surfaces (outer/middle for cortical layer II and III), D: RI values at the deep white matter surface of FA, ADC, apparent fiber density (AFD), and C: DWI connectome (DWIC) profile (i.e., edge strengths A_ij=1-500 sorted from the nearest to the farthest node). Three blocks of msResNet were separately pre-trained to extract three sets of deep feature vector k_i that are most predictive of C₁, C₂, and C₃. The resulting deep feature sets were combined in the framework of ensemble learning to predict three output probability values, [P(C₁: SOZ), P(C₂: IZ), P(C₃: non-EZ)] from the given input vector X_i. Note that the proposed ensemble learning neural network consists of three msResNet blocks that are pre-trained to perform three binary classifications, 1) C₁: SOZ vs. C₃: non-EZ, 2) C₁: SOZ vs. C₂: IZ, 3) C₂: IZ vs. C₃: non-EZ, in the framework of ensemble learning that can predict three classes more accurately in the output layer. Finally, a new MRI marker of the epileptogenic likelihood ε_i was estimated by non-linearly combining two epileptogenic potentials, seizure onset likelihood μ_i (i.e., P(C₁)) and spiking likelihood γ_i (i.e., P(C₂)), after factorized by two exponents, a₁ and a₂ that weigh the individual contribution of SOZ and IZ to overall epileptogenicity according to their degrees of MRI abnormalities. An attention layer¹⁰ was used to identify specific MRI modalities that play the most predictive role for correct classification. Nodes surgically resected iEEG-defined SOZ (or IZ) and preserved iEEG-defined non-SOZ (or non-IZ), in which all patients achieved long-term seizure freedom, were treated as the ground-truth EZ and non-EZ, respectively.

Results

The proposed ensemble network provided higher accuracy in correctly predicting three classes, C_1-3 in both model and validation cohorts (90% and 80%, Fig 2), compared with the standalone msResNet (86% and 72%, Fig 2). It also yielded a huge effect size of ε_i (i.e., Cohen’s d = 2.73) to differentiate the ground-truth EZ from non-EZ (Fig. 3A) and identify that reduced FA of cortical layers (II and III) and increased DWIC profile are the most predictive markers to localize the iEEG-defined SOZ and IZ (Fig. 3B).

Discussion

To the best of our knowledge, the present study is the first work reporting that a novel ensemble learning neural network of deep MRI features (intensity, diffusivity, connectivity) can predict the iEEG-defined SOZ and IZ more accurately and also identify the most predictive MRI markers that should be used to localize the EZ in current preoperative MRI protocol. This non-invasive prediction may help guide any type of invasive iEEG recording since it is solely based on the deep learning of preoperative MRI abnormalities, associated with iEEG abnormalities, that are often invisible on clinical MRIs. Furthermore, the proposed MRI-based epileptogenic likelihood can be used as a priori information to quantify the severity of seizure excitability.

Conclusion

The findings of this study suggest a promise of clinical MRI-based ensemble deep learning approach that can make specific iEEG abnormalities (SOZ and IZ) more readily appreciable via intelligently recognizing and highlighting highly discriminative but subtle MRI abnormalities associated with specific iEEG abnormalities. Further investigation will provide important implications for applying the proposed deep learning-based preoperative evaluation to clinical presurgical workups.

Acknowledgements

This research was supported by grants from the National Institutes of Health, R01 NS089659to J.J. and R01 NS064033 to E.A.