Abstracts

Rationale: Brain tissue classification, particularly identification of cortical gray matter (GM), is critical for many imaging and source localization methods used in patients with epilepsy. Widely used imaging analysis software packages provide excellent automated tissue classification in typical brains, but often perform suboptimally in structurally abnormal brains. Here we describe a template-free machine learning approach for cortical GM and white matter (WM) tissue classification based on multi-contrast multi-scale local MR image features. We compare our results to those obtained using several other widely available image processing software packages.

Methods: Study participants consisted of 3 groups: 10 healthy volunteers (HVs), 5 patients with malformations of cortical development (MCDs), and 10 patients with destructive brain lesions (LES) following surgery, stroke, or traumatic brain injury; all had 3T brain MRIs with 3D MPRAGE, T2, and FLAIR images. For each voxel, we compute 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. We created labelled training masks in 5 HVs using Freesurfer cortical GM and cerebral WM, and sMRIPrep-derived CSF. We trained a GM/WM/CSF classifier using multinomial logistic regression, then applied it to the remaining 5 HVs and all patients. For each group, the classification results were compared to those obtained using 1) Freesurfer (FS), 2) sMRIPrep (https://pypi.org/project/smriprep/) with FSL (sFSL), and 3) sMRIPrep with ANTs ATROPOS (sANTS). Classifications were compared qualitatively by visual inspection, and quantitatively using Dice Similarity Coefficients (DSC). Friedman test (Q) was used to assess similarity across methods.

Results: Quantitatively, our cortical GM and WM tissue classification yielded similar results across groups with significant heterogeneity across methods, appearing highly similar to Freesurfer (DSC GM 0.90-0.91, WM 0.91-0.92) and sFSL (DSC GM 0.0.87-0.90, WM 0.89-0.90), and less so to sANTS (DSC GM 0.72-0.80, WM 0.85) (across methods comparison, HV/MCD combined group: GM Q=15.2, p=0.001, WM Q=12.6, p=0.002; LES group GM Q=16.2, p=0.00; WM Q=9.6, p=0.008). Visually, compared to our classifier sANTS overestimates WM at the GM-WM junction, and overestimates GM at the GM-CSF junction. At the borders of destructive lesions, sFSL and our classifier best matched visual classification, with sFSL overestimating WM and ours GM; FS and sANTS often failed to classify these regions as brain tissue.

Conclusions: Our brain tissue classification method, FS, and sFSL perform similarly, while sANTS has more significant mismatches at tissue boundaries. In patients with destructive lesions, there were significant mismatches in structurally abnormal brain regions, with best results obtained using our classifier or sFSL. Our classification is atlas-free, requires only a small labelled training set, and can be easily applied to new patient populations with a variety of structural brain abnormalities.

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