Abstracts

Rationale:
Recently, artificial neural networks (ANNs) have been incorporated into the voxel-based morphometric analysis program (MAP18) to automate the detection of focal cortical dysplasia (FCD) in epilepsy patients (David et al., Epilepsia 2021). We performed external validation of the "chunkedNN" option for ANN in MAP18 with data that were naïve to the training set to determine its applicability to a new patient cohort.

Methods:
This retrospective study included patients with a 3D T1-weighted sequence, pathologically confirmed FCD II, and one year post-surgical follow-up. Age and gender matched healthy controls (HCs) were also included to assess false positive findings. MRI scans were acquired using 1.5T or 3T Siemens scanners for both patients and HCs. The majority of the 3D T1-weighted sequences had 0.8×0.8×1 mm voxel resolution and 256 slices. For each patient, a lesion label was created manually with MRIcron. Morphometric analysis with MAP18 on the T1-weighted images was performed using the “Large Average from All 1.5 and 3T Scanners (61 1.5 & 3T scanners with 3716 T1 images)” normal database to generate voxel-wise FCD probability maps in Montreal Neurological Institute (MNI) space using the ‘chunkedNN’ option (i.e., a cluster of ANNs trained on different randomized chunks of image data). Voxels were included in the prediction cluster if the probability was greater than 0.5. Voxels sharing a common face, edge, or corner were assigned a cluster; clusters with less than five voxels were discarded. Lesion labels were registered to MNI space with Statistical Parametric Mapping (SPM) in MATLAB for comparison with the clusters. Overlap was defined as the number of overlapping voxels divided by the number of voxels in the lesion label and the probability cluster. In patients, false positives (FPs) were defined by clusters outside of the lesion label; in HCs, FPs were defined by the presence of any cluster.

Results:
We included 100 patients (73 FCD IIB and 27 FCD IIA) and 42 HCs. Predicted clusters overlapped with at least one voxel of the lesion label in 56% (15/27) and 66% (48/73) of the FCD IIA and IIB groups, respectively (Figure 1A, Figure 2). For the FCD IIA patients with overlap, 89% (24/27) of them had overlap ≤ 0.2. In FCD IIB, where 88% (64/73) of the patients had overlap ≤ 0.2. At least one FP cluster was observed in 78% (21/27) of FCD IIA, and 63% (46/73) of FCD IIB; 48% (20/42) of HCs also had at least 1 FP cluster (Figure 1B). For FCD IIA, 51.3% (60/117) of FP clusters were ipsilateral to the lesion label; for FCD IIB, 48.3% (99/205) of FP clusters were ipsilateral. The HCs had fewer FPs than either patient cohort, with 76% (32/42) of HCs having fewer than three FPs, as compared to 48% (13/27) in FCD IIA and 67% (49/73) in FCD IIB.

Conclusions: The ‘chunkedNN’ ANN included in MAP18 shows efficacy in automatically detecting FCD II lesions in an external data set that was naïve to the training set, supporting its general applicability to a new patient cohort. Refinement is needed to further balance sensitivity and specificity.

Funding: Funding: NIH R01 NS109439