Abstracts

Rationale: Our recent study showed that postoperative fiber volume change in language pathways defined by the diffusion-weighted imaging-maximum a posteriori probability (DWI-MAP) method predicted the occurrence of a postoperative language deficit with sensitivity/specificity of 0.78+-0.05. In this method, however, the statistical measurement of fiber volume loss was often confounded by false-positive streamlines on DWI. To overcome this limitation, we propose a deep convolutional neural network (DCNN), which translates deep learning of electrical stimulation mapping (ESM)-based language pathways to further improve the quantification of postoperative fiber volume change. The present study of children with focal epilepsy investigated if the DWI-DCNN method would enhance the performance of prediction of a postoperative language score based on the postoperative changes of language-related fiber volume.  Methods: The clinical ESM procedure initially localized 14 types of eloquent areas supporting primary motor, language, and visual function in 118 patients on our database (age: 14.6+-7.1 years). The aggregate data yielded a probability map of each eloquent function delineated on a standard surface image (Fig. 1). This map then served as a cortical mask in standard template space to sort 14 functionally-important DWI pathways: C1-14, from 3T DWI streamline tractography, which included four different language pathways in the left hemisphere, C7-10: auditory expressive aphasia, visual expressive aphasia, auditory receptive aphasia, and speech arrest pathways. Our DCNN learned 3D coordinates of 100 equidistant tract segments in Ci=1-14. After training to minimize the center loss between Ci, the softmax layer produced the output probability vector. An argument of the maximum output probability was used to predict the class membership of a given tract, tj. This DCNN detected language pathways, C7-10, using pre- and postoperative DWI data of 17 children (age: 13.2+-3.1 years) who underwent pre- and postoperative language evaluation using language fundamental (CELF) tests. Postoperative tract volume change was defined by Di = (volume of preoperative Ci - the volume of postoperative Ci)/volume of preoperative Ci. Similarly, D7-10 was defined to quantify the postoperative change of expressive/receptive language, Δreceptive/expressive = (preoperative CELF score - postoperative CELF score)/preoperative CELF score. Canonical correlation analysis (CCA) determined the statistical significance of functional correlates between Di and Δreceptive/expressive.  Results: In training/testing sets, the proposed DCNN was converged at a center loss of 0.0008/0.0001, maximizing correct classification, F1 score of 0.986/0.987 across C1-14. The subsequent CCA analysis showed that postoperative volume changes of four language tracts, Δ7-10, determined by DWI-DCNN, were significantly correlated with the change of postoperative language outcome, Δreceptive/expressive (Fig. 2, r=0.55/0.67, p=0.02/0.00 for receptive/expressive), while no significant correlation was found between Δreceptive/expressive and Δ7-10 determined by DWI-MAP (r=0.42/0.40, p=0.09/0.11 for receptive/expressive). Conclusions: Our finding suggests that, compared with DWI-MAP analysis, the proposed DWI-DCNN may be more useful to localize the language pathways of which greater damage would lead to a more severe postoperative language deficit.  Funding: R01-NS089659 to J.W.J and R01-NS064033 to E.A