Abstracts

Rationale: Resective brain surgery is an effective but underutilized intervention for medically refractory focal epilepsy. The location and extent of resected brain regions relative to epileptogenic tissue contribute to postsurgical outcomes. Here, we trained 3D convolutional neural networks (CNNs) to label the resection cavity using postsurgical neuroanatomical MRI scans. We also labeled the hippocampal remnant and intact contralateral hippocampus in a subset of these cases that underwent temporal lobe resections.

Methods: Participants who underwent surgery for epilepsy treatment and had clinically acquired volumetric T1w MRI at NYU Langone hospital were included in the study (N = 568). Manual labeling was used to generate labels for CNN training and testing. Two labeling protocols were utilized to label (1) the surgical resection cavity for brain-wide focal epilepsy patients (N=433), and (2) hippocampal remnant, contralateral hippocampus and resection cavity in individuals who underwent temporal lobe resections (N=135). Brain-wide resection label dataset train/test split = 350/83; hippocampal remnant dataset train/test split = 89/46. Models were trained using a multi-scale 3D deep convolutional neural network approach implemented in the DeepMedic software package. Dice overlap and volumes of resection cavities and ipsilateral/remnant and contralateral hippocampi were assessed in independent testing cases.

Results: The CNN-based models successfully labeled the resection cavity and hippocampal remnant and the contralateral hippocampus in individuals who had temporal lobe resections. Mean resection volume = 26127 ± 13686 mm³, dice overlap = 0.78 ± 0.21. Hippocampal remnant volume = 795 ± 956 mm³, dice overlap = 0.59 ± 0.23. Contralateral hippocampal volume = 3636 ± 693 mm³, dice overlap = 0.81 ± 0.12. An example of labeled hippocampal remnant, contralateral hippocampus, resection cavity and ventricles in an out-of-sample case is shown in Figure 1. Figure 2 demonstrates longitudinal changes (1 day, 1 week and ~1 year) in the size and morphology of the resection cavity in an individual following surgical resection of a right frontal focal cortical dysplasia.

Conclusions: We have demonstrated that CNNs are well suited for labeling neuroanatomical changes following epilepsy surgery. Labeling is a fundamental processing step for a range of quantitative image analysis methods including volumetry, diffusion and functional MRI and network-based methods that utilize these imaging modalities. CNNs are an attractive approach for brains with gross abnormalities because they do not rely on template-based coregistration, which is a major limitation for the most widely used current morphometric techniques. CNN-based methods therefore enable quantitative neuroimaging-based assessment of postsurgical brains and may assist in presurgical planning and predicting health outcomes in individuals with severe epilepsy.

Funding: NIH R21 NS117990