Predictors of Epilepsy Presentation in Unruptured Brain Arteriovenous Malformations: A Quantitative Evaluation of Location and Radiomics Features on T2-Weighted Imaging

World Neurosurg. 2019 May:125:e1008-e1015. doi: 10.1016/j.wneu.2019.01.229. Epub 2019 Feb 13.

Abstract

Objective: To explore predictors of epilepsy presentation in unruptured brain arteriovenous malformations (bAVMs) with quantitative evaluation of location and radiomics features on T2-weighted imaging.

Methods: This retrospective study identified 117 patients with unruptured bAVMs. Cases were randomly split into training dataset (n = 90) and test dataset (n = 27). On the training dataset, we applied atlas-based analysis to identify epilepsy-susceptible brain regions of bAVMs, and then applied the radiomics technique to explore shape, intensity, and textural features that were correlated with epilepsy presentation. Informative radiomics predictors were selected by least absolute shrinkage and selection operator with 3-fold cross-validation. A linear classification score was then constructed, and we tested if we could precisely identify epilepsy-susceptible bAVMs with the location and radiomics predictors.

Results: Two brain regions and 4 radiomics features were screened out as predictors for epilepsy. The percent of damage of the right precentral gyrus and the right superior longitudinal fasciculus was associated with epilepsy presentation. The 4 radiomics features were Original_firstorder_Median, Wavelet-LHL_firstorder_InterquartileRange, Wavelet-HHL_firstorder_InterquartileRange, and Wavelet-HHH_glrlm_RunVariance. Epileptogenic bAVMs had larger variance of run lengths, larger median value, and interquartile range of voxel intensities. On the training dataset, these 6 predictors were able to classify epilepsy-susceptible bAVMs with accuracy at 0.822, and the area under the curve was 0.866 (95% confidence interval, 0.791-0.940). On the test dataset, sensitivity, specificity, and accuracy of classification reached 0.786, 0.769, and 0.778, respectively.

Conclusions: Epilepsy-susceptible bAVMs had distinct locations and radiomics features on T2-weighted imaging.

Keywords: Arteriovenous malformation; Classification; Epilepsy; Machine learning; Radiomics.

MeSH terms

  • Adult
  • Arteriovenous Fistula / complications
  • Arteriovenous Fistula / diagnostic imaging*
  • Brain / diagnostic imaging*
  • Brain / pathology
  • Brain / physiopathology
  • Epilepsy / complications
  • Epilepsy / diagnostic imaging*
  • Female
  • Humans
  • Image Processing, Computer-Assisted
  • Intracranial Arteriovenous Malformations / complications
  • Intracranial Arteriovenous Malformations / diagnostic imaging*
  • Magnetic Resonance Imaging
  • Male
  • Retrospective Studies
  • Sensitivity and Specificity
  • Young Adult