Neuroanatomic Markers of Posttraumatic Epilepsy Based on MR Imaging and Machine Learning

Abstract

Background and purpose: Although posttraumatic epilepsy is a common complication of traumatic brain injury, the relationship between these conditions is unclear and early posttraumatic epilepsy detection and prevention remain major unmet clinical challenges. This study aimed to identify imaging biomarkers that predict posttraumatic epilepsy among survivors of traumatic brain injury on the basis of an MR imaging data set.

Materials and methods: We performed tensor-based morphometry to analyze brain-shape changes associated with traumatic brain injury and to derive imaging features for statistical group comparison. Additionally, machine learning was used to identify structural anomalies associated with brain lesions. Automatically generated brain lesion maps were used to identify brain regions where lesion load may indicate an increased incidence of posttraumatic epilepsy. We used 138 non-posttraumatic epilepsy subjects for training the machine learning method. Validation of lesion delineation was performed on 15 subjects. Group analysis of the relationship between traumatic brain injury and posttraumatic epilepsy was performed on an independent set of 74 subjects (37 subjects with and 37 randomly selected subjects without epilepsy).

Results: We observed significant F-statistics related to tensor-based morphometry analysis at voxels close to the pial surface, which may indicate group differences in the locations of edema, hematoma, or hemorrhage. The results of the F-test on lesion data showed significant differences between groups in both the left and right temporal lobes. We also saw significant differences in the right occipital lobe and cerebellum.

Conclusions: Statistical analysis suggests that lesions in the temporal lobes, cerebellum, and the right occipital lobe are associated with an increased posttraumatic epilepsy incidence.

FIG 1.

The VAE network and an input/output sample pair from the ISLES data set. X denotes the input data; Z denotes its low-dimensional latent representation. The VAE consists of an encoder network that computes an approximate posterior q_φ (Z|X), and a decoder network that computes p_θ (X|Z). The VAE model takes T1, T2, and FLAIR images from individual subjects (left), compresses them to generate a latent representation (Z), and regenerates 3 images (right). The VAE is trained on a data set that contains few lesions. After training, when presented with a newly lesioned brain, the reconstruction effectively removes the lesion from the image, resulting in a normal (lesion-free) version of the brain.

FIG 2.

Three orthogonal views through the P values thresholded at P = .05 (FDR-corrected) obtained for the F-test for TBM analysis using Jacobian determinants. L indicates left; R, right.

FIG 3.

Reconstruction results obtained by applying the VAE to the ISLES data set. A, Sample slices from input images. B, Slices reconstructed from the VAE. C, Difference between input and reconstructed images. D, Error maps after applying median filtering to reduce the occurrence of spurious voxels. E, Manually delineated lesion masks used as ground truth to evaluate VAE performance.

FIG 4.

Orthogonal views through the P values thresholded at P = .05 (FDR-corrected) obtained for the F-test comparing lesion maps for the PTE and non-PTE groups. L indicates left; R, right.