Article
Diffusion MRI metrics-based aphasia prediction based on tract orientation mappings
Diffusions-MRI-Metriken basierte Vorhersage von Aphasie in Bezug auf Faserbündel-Orientierungs-Karten
Search Medline for
Authors
-
Boshra Shams
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland; Humboldt Universität Berlin, Cluster of Excellence: „Matters of Activity. Image Space Material“, Berlin, Deutschland
-
Klara Reisch
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
-
Peter Vajkoczy
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland
-
Thomas Picht
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland; Humboldt Universität Berlin, Cluster of Excellence: „Matters of Activity. Image Space Material“, Berlin, Deutschland
-
Lucius Fekonja
- Charité – Universitätsmedizin Berlin, Klinik für Neurochirurgie, Berlin, Deutschland; Humboldt Universität Berlin, Cluster of Excellence: „Matters of Activity. Image Space Material“, Berlin, Deutschland
| Published: | May 25, 2022 |
|---|
Outline
Text
Objective: Damage to the major white matter (WM) structure caused by brain tumors may lead to various degrees of dysfunction. The aim of this study was to predict language impairments (aphasia) in brain tumors invading language networks. To this end, we applied machine-learning based analysis using dMRI-based metrics within regions related to language pathways using automated tractography.
Methods: 80 patients with left-hemispheric language-eloquent gliomas (WHO grade II (10), III (24), IV (44)) and metastases (2) were included (32 females, 48 males, average age=52.71, SD=14.4, age range 21-83). The language status was assessed preoperatively using Berlin Aphasia Score (BAS), labelling non-aphasics with 0 (BAS=0, n=55) and aphasics with 1 (BAS>0, n=25). Tract Orientation Mappings (TOMs) related to language function in both hemispheres were generated automatically in individual space using TractSeg. DMRI metrics (AD, ADC, FA and RD) were sampled within specific tracts in the left hemispheres using TOMs. Z-score normalization per voxel was performed based on values of the corresponding tracts in the right hemisphere and mean and std were extracted as features, subsequently. A support vector machine (SVM) with RBF kernel and random forest-based feature selection method was trained using dMRI-related WM statistics (mean and std per bundle) and was further evaluated with 5-fold cross-validation to predict aphasia.
Results: The best model performance was achieved with features extracted from regions related to language pathways based on dMRI metrics reaching up to accuracy 77.7% (specificity 79.4%, sensitivity 74% and AUC 76.7%). Features from AF (arcuate fasciculus), MLF (medial longitudinal fasciculus) and IFOF (inferior fronto-occipital fasciculus) ROIs were selected as the most effective features and RD and AD were the most effective dMRI metrics, respectively. Moreover, by incorporating age, gender and tumor WHO grade with dMRI-based features, the model performance remained the same with only a slight change in specificity (78.5%) and sensitivity (75.8%).
Conclusion: We have demonstrated that the detailed WM analysis of individual patients using various dMRI-based features allows us to predict language impairment with high accuracy. Also, this study confirmed that AF is the main pathway to maintain language function. The model’s high accuracy points towards a relevant potential to enhance future preoperative neurosurgical planning and analyses.
