Abstracts

Rationale: Traumatic brain injury (TBI) can produce various intracranial abnormalities, such as contusions and hemorrhages (Epilepsia 44, 2003: 11-17). Up to 50% of patients with TBI develop posttraumatic epilepsy (PTE). PTE manifests with recurrent, unprovoked seizures at least one week after TBI and is preceded by a latent period during which clinical intervention may prevent epileptogenesis (Neurobiology of Disease 123, 2019: 115-121). One recent challenge is to validate biomarkers of epileptogenesis, which is the main aim of the Epilepsy Bioinformatics Study for Antiepileptogenic Therapy (EpiBioS4Rx). Lesion segmentation makes it possible to identify the locations and extent of intracranial alterations and can be useful to investigate the pathophysiology of PTE and its potential relationship with specific lesion characteristics. We manually segmented lesions in TBI patients from the EpiBioS4Rx cohort to evaluate if lesion volume is related to seizure development and to injury severity. Currently, automatic segmentation algorithms optimized for TBI patients do not exist. Thus, this study is also motivated by the need for manual lesion segmentations to establish ground truths to train and implement automatic, machine learning-based methods to identify brain lesions due to a TBI. Methods: Lesions of 24 TBI patients from the EpiBioS4Rx cohort were segmented. These patients were divided into two groups based on if they experienced at least one seizure after the TBI. Demographic information is reported in Table 1. Parenchymal lesions were manually segmented with ITK-SNAP (Neuroimage 2006 Jul 1; vol. 31 no. 3: 1116-28) from T2-weighted fluid attenuated inversion recovery (FLAIR) images. When lesion characterization was uncertain, corresponding T1-weighted Magnetization Prepared Rapid Acquisition Gradient Echo (MPRAGE) images were used to identify the lesions. In a subset of patients, it was possible to clearly distinguish the lesion core from the surrounding edema: in those cases, two different segmentation masks were obtained, one per each component (edema and lesion). We assessed with the Wilcoxon rank-sum test if the total lesion volume, the lesion core volume, and the edema volume are statistically different for the epileptic and not epileptic groups. Finally, we used the Kruskal-Wallis test to assess the statistical association between the lesion volumes and the degree of severity measured with the Glasgow Coma Scale (GCS). Results: We found 13 TBI patients with lesions in the frontal lobe, 6 in the parietal lobe, 3 in the occipital lobe, and 10 in the temporal lobe. Some patients had lesions in multiple lobes. In this cohort, we did not find any association between lesion position and seizure development or lesion volume and GCS. However, total lesion volumes proved to be statistical different for the two groups (p-value < 5.392*10-7) as well as the lesion core volumes (p-value < 2.455*10-6) and the edema volumes (p-value < 0.0004341). Figure 1 shows a manual segmentation for an axial brain slice (left) and the total lesion volume distribution for the epileptic and not epileptic groups (right). Conclusions: Manual segmentation of brain lesions in TBI patients can be very helpful to gain more insight into the mechanisms underlying seizure development. Indeed, we found that lesion volume is statistically related to seizure occurrence with a p-value < 5.392*10-7. EpiBioS4Rx is an ongoing study that will enroll 300 patients, so we will obtain conclusive results and additional segmentations to use as ground truth to develop automatic segmentation systems. Funding: This study was conducted with the support of the National Institute of Neurological Disorders and Stroke (NINDS) of the National Institutes of Health (NIH) under award number U54 NS100064 (EpiBioS4Rx).