Abstracts

Rationale: The gold standard for identification of the epileptogenic zone (EZ) is the area of resection in patients who become seizure-free postoperatively. Therefore, creation of resection masks using postoperative MR images is critical to defining the EZ. Manual resection mask creation is labor intensive and prone to error. However, accurate semi/fully automated resection mask creation is reliant on accurate brain tissue classification and registration of pre- and post-op MR scans. While tissue classification using available brain imaging analysis packages is accurate in typical brains, classification accuracy varies widely in post-resection brains. In this study we use a template-free machine learning approach to brain tissue classification based on multi-contrast multi-scale local image features. We compare the obtained pre- to post-op difference images to those obtained using sMRIPrep, a freely available optimized imaging pipeline, as well as to manually optimized resection masks.

Methods: Study participants consisted of 10 patients with non- or subtly lesional epilepsy and 5 healthy volunteers (HVs) with available pre- and post-operative 3T MR images that included 3D T1, T2, and FLAIR images. For each voxel in each image, we computed 3D rotationally invariant Gaussian, gradient magnitude, Laplacian and Hessian magnitude filters at 3 spatial scales (FWHM 2, 4, 8 mm). Images were co-registered within each subject, yielding a 39 feature signature for each voxel. To train our classifier, we created labeled training masks in 5 HVs using Freesurfer cortical gray matter (GM) and cerebral white matter (WM), and sMRIPrep-derived CSF, which we used in a multinomial logistic regression model. We applied this trained model to the pre- and post-operative images in our patients using a winner-take-all classification system. The same images were run through the standardized sMRIPrep tissue classification pipeline (https://pypi.org/project/smriprep/). Pre- and post-op T1 images were co-registered along with each tissue classification. The combined gray and white matter masks were subtracted to obtain a pre- versus post-op difference image. For each patient, the 2 difference images were compared visually, and compared quantitatively to a manually optimized resection mask using Dice Similarity Coefficients (DSC).

Results: On visual inspection, the most significant difference between the difference images was an underestimation of cortical GM and overestimation of CSF on the post-op image using the sMRIPrep brain tissue classification, with a resulting expansion of the resection area. Compared to the manually drawn resection masks, DSC showed low similarities for both images (DSC 0.28 for our classifier and 0.09 for sMRIPrep, and a DSC of 0.23 between our classifier and sMRIPrep), driven largely by the different estimates of diffuse brain volume loss.

Conclusions: Our template-free machine learning approach to brain tissue classification facilitates automated resection mask creation compared to the FSL-based segmentation used in sMRIPrep, driven primarily by improved tissue classification in the post-resection brain images.

Funding: Please list any funding that was received in support of this abstract.: Funding was provided through the NIH Intramural Research Program.