Probabilistic Presurgical Language fMRI Atlas of Patients with Brain Tumors

DISCUSSION

Language atlases provide templates for quantitative assessment of language mapping, such as localization, lateralization, and strength, as well as for automated detection of language networks with resting-state fMRI.^5⇓-7 This study presents the first language atlas built from preoperative fMRI results of patients with brain tumors using the language paradigms similar to those recommended by the American Society of Functional Neuroradiology (ie, SENT and Silent Word Generation).¹⁵

Although patients with brain tumors can have discernible intersubject variation in functional activations, language atlases from this study involve many fMRI activations projected onto a standard space. Thus, we were able to reproduce known characteristics of the language network. For example, critical language regions that are commonly assessed clinically, such as the Broca area (aPLA), Wernicke area (pPLA_T), angular and supramarginal gyrus (pPLA_P), DLPFC, pre-SMA, and BTLA, can be observed.^18,27,28 Left-hemisphere dominance is still observed considering the higher probability of overlap in left-hemisphere PLAs and ancillary language areas compared with their right-hemisphere counterparts.^27,28

In agreement with the literature, we observed that the semantic task (SENT) was more likely to activate posterior PLAs compared with the 2 Silent Word Generation tasks (CAT and LETT).^15,29 This observation can be attributed to sentence-completion tasks being more proficient at activating the posterior language network.^29,30 Across all 3 paradigms, the higher probability of overlap in pPLA_T compared with pPLA_P agrees with a previous study of language regions in presurgical fMRI.¹⁸ We also found that the probability of detecting activation in the right DLPFC for the CAT task was higher than for SENT and LETT tasks. This finding can be attributed to more likely recruitment of the right DLPFC for the CAT task, which is corroborated by a separate study on verbal fluency paradigms,³¹ and to the involvement of the right DLPFC in retrieval tasks.³² The probability of detecting activation in the SMA for both hemispheres was similar across the 3 paradigms, a likely consequence of the intersubject variation during spatial normalization and the 6-mm isotropic smoothing applied during preprocessing. Considering that SMA laterality is known to corroborate language laterality,⁵ it can still be observed that the spatial extent of SMA activations is asymmetric toward the left hemisphere on the POM.

The pooled probabilistic language atlas developed in this study is equivalent to the weighted average of the 3 language paradigm POMs. Thus, the pooled probabilistic language atlas developed in this study emphasizes areas of activation common across the 3 paradigms. On the basis of use case and context, the pooling strategy and choice of paradigms to include can differ. Thus, we had also created language atlases from single paradigms and made them available. The distinguishing attribute between our language atlases and others developed from healthy individuals is information on intersubject variation. In our atlases, this difference is encapsulated by the spatial extent of the atlas at a specified overlap threshold, with higher thresholds corresponding to a lower tolerance for variation. For example, the spatial extent of areas with higher overlap in the frontal lobe is continuous between the proximally close aPLA and DLPFC in the left hemisphere and can be attributed to using a population of patients with brain tumors for atlas building; these patient populations are heterogeneous in terms of functional anatomy.⁹ Given the implementation of the POM methodology, sources of intersubject variation taken into consideration include intrinsic variation, normalization imprecision due to lesion-distorted anatomy, and functional reorganization due to tumor invasion.

We compared our results with those of Neurosynth because it is a widely-referenced large-scale platform for automated synthesis of fMRI data. The Neurosynth result is a statistical inference map generated using a χ² test of independence and informs if coordinates of a voxel have been reported more consistently in studies involving the term “language” than in studies that did not.²⁰ Therefore, it allowed us to calculate Dice coefficients in a similar fashion among varying thresholds to ensure a fair comparison. The comparison demonstrated that our templates had better agreement with individual patient’s fMRIs. This outcome may be attributed to the Neurosynth meta-analysis including fMRI studies with different cohorts, primarily healthy individuals, various task and resting-state fMRI paradigms, and different acquisition and analysis methods, whereas our atlas was built from a uniform source of fMRI data. In addition, for patients with brain tumors, the position of a tumor with respect to the classic/principal functional anterior Broca and posterior Wernicke areas can influence the locoregional functional reorganization. Given the large group of patients included in the study, it is possible that filtering for contributing patients with similar tumor locations (eg, anterior or posterior) during atlas development may allow better spatial agreement with the evaluation data set.

We envision our atlas as having potential clinical and research applications through the derivation of templates for ROIs. An example would be language assessment of patients with brain tumors in which templates of functional ROI atlases based on healthy individuals are often used to calculate the laterality index. Templates generated from our atlas account for functional anatomy distortions due to the tumor, which may improve the fidelity of the calculated laterality index.^5,33 Another potential application is in imaging biomarker studies, wherein postwarping of primary and ancillary language areas into patient space, templates from our atlas can serve as language localizers for quantifying imaging measures, whether from fMRI activations of more specific paradigms, or different MRI sequences, or different imaging modalities.^34,35 Our atlas could also be used to guide resting-state fMRI template matching for detecting language networks.^6,36 These applications could rely on manual delineation of language areas by experts, but atlases help to make the process automated and less operator-dependent. The potential clinical implications of this study will depend on the improvement with each different use of the atlas, eg, more accurate assessment of language lateralization, imaging biomarker quantification, and resting-state fMRI language mapping. Further validation studies and clinical trials may be needed to assess the practicality and effectiveness of the atlas in clinical settings. Although the existence of large lesions and distorted anatomy may introduce the potential inaccuracy of spatial registration,³⁷ recent studies have demonstrated that such errors are reduced with more modern deformable registration methods, allowing reasonably good performance even in patients with brain tumors.³⁸ In addition, for large ROIs such as PLAs, the misregistration may remain local and its effect would be specific to applications (eg, laterality calculation, biomarker quantification, template matching). If the ROI is far from the tumor, eg, using the aPLA atlas to calculate the laterality index for a patient’s tumor near the pPLA, the effect of spatial registration should be minimum. If the use of an atlas directly focuses on the language ROI covering/adjacent to lesions, we would recommend that the users inspect the registration closely. Again, because the ROIs have large spatial extents, the overall effect may not be significant. However, when there are concerns, one can consider extending the ROIs by applying a lower probability threshold on the atlas. This is one of the advantages of the probabilistic atlas.

This study has some limitations. First, this is a single-institution study with MRI data gathered from limited scanner platforms. The data set reflects the typical patient population undergoing presurgical fMRI in our institution. Including data from other institutions, which would incorporate more diverse patient cohorts and/or fMRI paradigms/methods, could further refine the atlas. In parallel, multicenter studies have shown that fMRI data acquired on different scanners will have different activation effect sizes and spatial smoothness.³⁹ This result may impact the atlas on areas with lower overlap probabilities (extents). However, we expect the central tendencies of high-overlap probabilities to remain similar, and corresponding brain regions identified in our atlas would still be useful for appropriate language-related analyses. Second, our atlas was developed using only data from patients with brain tumors. However, a substantial portion of presurgical fMRI studies are performed for patients requiring brain tumor resection.⁴⁰ Third, only clinical language generation (LETT and CAT) and semantic paradigms (SENT) were used to develop the atlas. Thus, the functional anatomy typically recruited in these tasks is emphasized in our atlas. Other fMRI task paradigms may yield different weightings in language areas and may involve additional brain regions. The potential implications of this variation would likely depend on its application and whether the weights and the additional regions are used, eg, for assessing language lateralization or for assisting resting-state fMRI analysis.