Authors :
Presenting Author: Min-Hee Lee, PhD – Wayne State University
Masaki Sonoda, MD, PhD – Assistant Professor, Neurosurgery, Yokohama City University; Csaba Juhász, MD, PhD – Professor, Pediatrics, Wayne State University; Eishi Asano, MD, PhD – Professor, Pediatrics, Wayne State University; Jeong-Won Jeong, PhD – Professor, Pediatrics, Wayne State University
Rationale:
Accurate prediction of postsurgical language improvement may encourage resective surgery in children with drug-resistant epilepsy. Diffusion weighted imaging connectome (DWIC) has been used to identify white matter biomarkers associated with language functions. However, low angular resolution of tractography acquisition inevitably creates false-positive tracts, limiting the reliability of the identified biomarkers. This study proposes a deep convolutional neural network (DCNN) to effectively control the false-positive tracts in tractography data. Its promise was validated by demonstrating how much it can improve the predictability of postsurgical language outcomes using presurgical DWIC data of children who undergo resective surgery.
Methods:
Retrospective data (i.e., presurgical tractography, pre and postsurgical neuropsychological language tests performed before and after surgery) were obtained from 40 children (11.9±4.0 years old, 19 boys) with drug-resistant epilepsy who underwent two-stage surgery. Whole-brain backbone DWIC was constructed based on the AAL node atlas (https://www.gin.cnrs.fr/en/tools/aal/) with 1477 true positive (or reference) tract classes extracted from whole-brain tracts. Briefly, the DCNN model was trained to learn 3-D coordinates of each reference tract by minimizing center loss (Fig. 1A). Three DWICs were constructed from the presurgical whole brain tractography data including 1) raw DWIC including the whole-brain tracts, 2) backbone DWIC without DCNN, and 3) backbone DWIC with DCNN (Fig. 1B). At each DWIC, the nodes of language-related connectome (LRC) were identified by measuring significant correlations (p< 0.05 after Bonferroni correction) between the number of tracts connecting two nodes and presurgical language scores. Finally, four LRC markers: regional efficiency (RE), betweenness centrality (BC), information centrality (IC), current flow (CF) were extracted from individual nodes, fused using supervised multivariate canonical correlation, and analyzed using a multilayer perceptron to predict the presence of postsurgical language improvement. The perceptron model was trained in the model cohort (n=30, including 16 with improvement in neuropsychological testing). The balanced accuracy (BA) of the trained model was evaluated in the validation cohort (n=10, including 5 with improvement).