Supplement: Multicenter validated detection of focal cortical dysplasia using deep learning

Objective. To test the hypothesis that a multicenter-validated computer deep learning algorithm detects MRI-negative focal cortical dysplasia (FCD).

Methods. We used clinically acquired 3D T1-weighted and 3D FLAIR MRI of 148 patients (median age, 23 years [range, 2-55]; 47% female) with histologically verified FCD at nine centers to train a deep convolutional neural network (CNN) classifier. Images were initially deemed as MRI-negative in 51% of cases, in whom intracranial EEG determined the focus. For risk stratification, the CNN incorporated Bayesian uncertainty estimation as a measure of confidence. To evaluate performance, detection maps were compared to expert FCD manual labels. We also tested sensitivity in an independent cohort of 23 FCD cases (13±10 years). Applying the algorithm to 38 healthy and 63 temporal lobe epilepsy disease controls tested specificity.

Results. Overall sensitivity was 93% (137/148 FCD detected) using a leave-one-site-out cross-validation, with an a..., 1. Description of methods used for collection/generation of data [noel_deepFCD_patch_16x16x16.h5]:
To create the HDF5 dataset, for each of the 148 FCD patients, we sampled at most 1,000 cortical patches (or # voxels in the lesion, whichever is lower) of size 16×16×16 within the lesion on pre-processed T1- and T2-weighted FLAIR MRI. The same number of cortical patches were sampled randomly outside the lesion. The resulting lesional and non-lesional patches were concatenated, shuffled (to add another layer of randomization), and saved along with their binary labels (as a compressed HDF5 dataset). Refer original publication and FCD_Detection_Neurology_Supplement.docx for more details

2. Methods for processing the data [noel_deepFCD_patch_16x16x16.h5]:
MRI pre-processing involved linear registration to the MNI152 symmetric template, non-uniformity correction, intensity standardization with scaling of values between 0 and 100, and skull-stripping using an in-house deep learning method (..., FCD_Detection_Neurology_Supplement.docx contains:

Table e-1. MRI acquisition parameters across sites

Figure e-1. Hierarchical patch-based feature learning using CNN

Additional Methods:

Classifier design

Source code and data availability

Table e-2. Peak location of FCD lesions in MNI space

Table e-3. Peak location of false positive clusters in MNI space

eReferences

noel_deepFCD_patch_16x16x16.h5 (size: 6.4GB) contains two variables (data and labels):

also available from: https://doi.org/10.5281/zenodo.3239446

variables
array shape
description

data
{282736, 2, 16, 16, 16}
The data variable contains a numpy array (numpy.float) with 2,82,736 multimodal (T1- and T2-weighted FLAIR) patches of size 16×16×16. See Source code and data availability section in Additional Methods (FCD_Detection_Neurology_Supplement.docx) for details

labels
{282736}
...