gms | German Medical Science

Objective: Reconstruction of white matter tract profiles and quantifying diffusion measures along them based on tractography have been widely used to study local microstructural changes caused by tumor effects (e.g edema or infiltration). This study sought to investigate how the changes in different dMRI-based metrics could be associated with and predictive of motor status in patients with brain tumors in motor area using machine-learning based analysis.

Methods: Constrained spherical deconvolution (CSD) based probabilistic tractography was performed for corticospinal tract (5000 streamlines) on 116 motor area glioma II°-IV° patients (43 females, 73 males, average age=48.24, SD=16.47, age range 20-78) with affected preoperative motor status as 1 (MRC grade <5, n=45) and unaffected one as 0 (MRC grade 5, n=71). Along tract profile was generated in individual space per hemisphere (healthy and pathologic) for DTI-based different diffusion measures (FA, ADC, RD and AD) and fixel-based apparent fiber density metric (FD). These features correspond to tract profile statistics (mean, std, kurtosis and skewness) per each measure for both hemispheres. A machine learning model based on SVM using linear kernel was trained on tractometry-extracted features, gender, age and tumor type and further evaluated with leave-one-out cross-validation to predict motor status.

Results: The best model performance was achieved with features extracted from FD, FA, RD, AD and ADC measures reaching up to accuracy 74.13% (specificity 73.23% and sensitivity 75.5%). Additionally, our model was trained with i) FD, FA, ADC and ii) FD, FA, RD, AD set of features separately, and showed the accuracy level of 66.37% (specificity 67.6% and sensitivity 64.44%) and 74.13% (specificity 74.64% and sensitivity 73.33), respectively. It has been shown that demographic features can be derived from WM structure. However, such features and tumor type didn’t improve the performance of the model as the information has already existed in WM structure and even tended to bias the results (accuracy 70.68, Specificity 73.23% and sensitivity 66.6%).

Conclusion: Detailed analysis of corticospinal tract characteristics applying tractometry-extracted features and a novel machine-learning based model allows to associate motor status and individual white matter integrity with high accuracy, potentially improving surgical planning in the future.