Introduction

Recent meta-analysis¹ indicates that children with drug-resistant focal epilepsy (FE) should be referred for presurgical assessment in a timely manner. Early surgery can facilitate a more robust improvement of language function in subsets of FE children. This study investigates if an advanced imaging approach combining deep convolution neural network (DCNN)-based tract classification^2-4 with diffusion-weighted imaging connectome (DWIC) can serve as accurate preoperative imaging markers capable of predicting postoperative language outcome.

Methods

Pre- and postoperative DWI data of 25 FE patients (age: 10.21±2.5 years) were collected using a 3T MRI scanner with 55 isotropic gradient directions and b0 = 1000 s/mm³. For each patient, whole-brain tractography was reconstructed by tracking from 50 seeds in each white-matter-masked voxel and modeling diffusion peak orientations using MRtrix package (http://www.mrtrix.org/). AAL parcellation atlas (116 nodes, http://www.gin.cnrs.fr/en/tools/aal-aal2/) was used to create an adjacency matrix, S(u,v), in which each element defined a connection class, C_i, consisting of whole-brain total count of streamlines: f_n, connecting both u^th and v^th nodes. To standardize a whole-brain backbone network⁵, B(u,v) from a given matrix, S(u,v), 1477 true positive classes of S(u,v) were identified by DCNN-DWIC tract classification^2-4 (Fig. 1), where each tract of S(u,v): f_n(x,y,z) was automatically classified into one of the 1477 true positive classes, C_i=1-1477, predefined for B(u,v). To construct a modular language network significantly associated with postoperative improvement of receptive language (Fig. 2A), functional correlation was performed between postoperative changes of 1) CELF⁶-based receptive language score: Δ_NP = [score_post – score_pre] and 2) tract count of B(u,v): Δ_DWIC = [tract count in B(u,v)_post – tract count in B(u,v)_pre]. At each node of the modular network, Cohen’s d-value of preoperative local efficiency (LE)⁷ was evaluated using d-value = (m-μ)/s, indicating the distance between the LE value of the patient (m) and the mean of the LE values in the healthy control group (μ) in units of the standard deviation (s) of the LE values in the healthy control group. Finally, one-way ANCOVA (factor: Δ_NP ≥ 0 vs. Δ_NP < 0; dependent marker: Cohen’s d-value of preoperative LE, 5 covariates measured before surgery: age, seizure frequency, history of secondary generalized seizures, history of status epilepticus, total number of antiepileptic drugs tried) was performed to identify the nodes of which LE values were within or greater than the range of the control group, supporting a high likelihood of language improvement after surgery. The LE values of the identified nodes were used as new imaging markers for postoperative receptive language improvement.

Results

Our modular network approach based on the DCNN-DWIC tract classification revealed two nodes in preoperative DWIC data, including the left middle temporal gyrus (L.MTG) and left angular gyrus (L.ANG), of which preoperative LE values are not significantly different in FE patients having Δ_NP ≥ 0 (i.e., postoperative improvement/preservation of receptive language) compared with the age-matched healthy controls (Fig. 2B). These measures may represent potential markers predictive of postoperative improvement in receptive language function based on the preoperative DWIC data. That is, these preoperative d-values of LE significantly differed between Δ_NP ≥ 0 and Δ_NP < 0 (Fig. 3A), yielding a significant correlation between preoperative d-value and DNP (Fig. 3B). The subsequent binary regression analysis using d-values as predictors achieved high accuracy (92.9±4.5%) to predict postoperative improvement in random permutation test. This outstanding accuracy was achieved by our DCNN-DWIC tract classification significantly enhancing the group difference, Δ_NP ≥ 0 vs. Δ_NP < 0 in preoperative LE (rated by one-way ANOVA F-statistics) by 398% (F=2.05/8.14 without/with DCNN) and 259% (F=1.45/3.76 without/with DCNN) in two language hubs: L.MTG and L.ANG, respectively (Fig. 4).

Discussion

Our findings provide new mechanistic insights as to whether the preservation (or increase) of local efficiency in a preoperative modular network predicts postoperative language improvement. Although preliminary, our data directly inform the clinical value of the proposed DCNN-DWIC analysis for accurate prediction of postoperative language improvement prior to resection, which may be expanded for prediction of other neurocognitive improvements such as full-scale IQ and fine motor skills.

Conclusion

This study proposes a new DWIC approach using DCNN tract classification as an effective imaging tool for the prediction of postoperative language improvement. Our new imaging markers, at the level of specific modular networks, could be useful to identify the neural correlates and potential predictors of long-term, specific, neurocognitive consequences associated with surgical intervention.

Acknowledgements

This work was supported by grants from the National Institutes of Health (R01 NS089659 to J.J. and R01 NS064033 to E.A.).

References

[1] Pestana Knight EM, Schitz NK, Bakai PM, Korokian SM, Lhatoo SD, Kaiboriboon K. Increasing utilization of pediatric epilepsy surgery in the United States between 1997 and 2009. Epilepsia 2015;56:375-81. [2] Xu H, Dong M, Lee MH, O’Hara N, Asano E, Jeong JW. Objective detection of eloquent axonal pathways to minimize postoperative deficits in pediatric epilepsy surgery using diffusion tractography and convolutional neural networks. IEEE Trans Med Imaging 2019;38:1910-22. [3] Lee MH, O’Hara N, Jeong JW. Improving reproducibility connectome analysis using deep convolutional neural network model. Proc Intl Soc Mag Reson Med. 27: 3578, Montreal, Canada, May 11-16, 2019. [4] Jeong JW, Lee MH, Juhasz C, Asano E. Deep learning DWI connectome improves detection of language network alteration in epilepsy children. Annual Meeting of the Organization for Human Brain Mapping, 25: 2382, Rome, Italy, June 9-13, 2019. [5] Gong G, He Y, Concha L, Lebel C, Gross DW, Evans AC, Beaulieu C. Mapping anatomical connectivity patterns of human cerebral cortex using in vivo diffusion tensor imaging tractography. Cereb Cortex 2009;19:524-36. [6] Semel E, Wiig EH, Secord WA. Clinical evaluation of language fundamentals, fourth edition (CELF-4). Toronto, Canada: The Psychological Corporation/A Harcourt Assessment Company, 2003. [7] Rubinov M, Sporns O. Complex network measures of brain connectivity: uses and interpretations. Neuroimage 2010;52:1059-69.