Abstract
BACKGROUND: Intersubject variation among patients with brain tumors complicates the template matching process for detecting the resting-state (rs) functional MRI (fMRI) language network when using independent component analysis (ICA). PURPOSE: This study aimed to develop methods that use a probabilistic language atlas to incorporate intersubject variation in brain tumor patients into the template matching process. METHODS: This retrospective study included 79 patients with brain tumors (average age, 50 ± 15 years) who underwent presurgical task-based (tb)-fMRI and rs-fMRI at clinical 3T scanners. At varying template generation thresholds (τ), binary and probabilistic templates were obtained from the language atlas. A binary template developed using healthy individuals and applied in published studies was tested as a control. Binary templates with goodness-of-fit (GOF) and probabilistic templates with weighted GOF (wGOF) or Jensen-Shannon distance (JSD) were used for template matching to recommend language networks from ICA. Dice coefficient and Pearson correlation with respect to tb-fMRI were assessed. Qualitative evaluation was performed by two neuroradiologists. Significant differences in Dice coefficients and Pearson correlations between different template matching methods, and significant differences between ratings were determined with the Wilcoxon signed-rank test, with p < 0.05 indicating statistical significance. RESULTS: Compared to the control, recommendations from methods involving probabilistic language templates had significantly higher (p < 0.05) Dice coefficients across τ = 0% to τ = 35%. Pearson correlation results followed similar trends to those for Dice coefficient. Dice coefficients with the control were found to be 0.247 on average. Peak Dice coefficients with using wGOF and JSD for template matching were found to be 0.349 and 0.350, on average, respectively. Likert score results indicated significantly superior performance (p < 0.05) in identifying ICA components that contain the language network by using probabilistic language templates. 58 and 52 of 79 recommendations for the control were found to contain the language network (Likert scores ≥4) based on rater 1 and 2, respectively, which increased to 69-73 with template matching using wGOF and JSD. CONCLUSIONS: The proposed application of probabilistic templates derived from a language tb-fMRI atlas of patients with brain tumors and probabilistic template matching methods can improve detection of rs-fMRI ICA language network in brain tumor patients.