A quantitative comparison of atlas parcellations on the human superior temporal sulcus

The superior temporal sulcus (STS) has a functional topography that has been difficult to characterize through traditional approaches. Automated atlas parcellations may be one solution while also being beneficial for both dimensional reduction and standardizing regions of interest, but they yield very different boundary definitions along the STS. Here we evaluate how well machine learning classifiers can correctly identify six social cognitive tasks from STS activation patterns dimensionally reduced using four popular atlases (Glasser et al., 2016; Gordon et al., 2016; Power et al., 2011 as projected onto the surface by Arslan et al., 2018; Schaefer et al., 2018). Functional data was summarized within each STS parcel in one of four ways, then subjected to leave-one-subject-out cross-validation SVM classification. We found that the classifiers could readily label conditions when data was parcellated using any of the four atlases, evidence that dimensional reduction to parcels did not compromise functional fingerprints. Mean activation for the social conditions was the most effective metric for classification in the right STS, whereas all the metrics classified equally well in the left STS. Interestingly, even atlases constructed from random parcellation schemes (null atlases) classified the conditions with high accuracy. We therefore conclude that the complex activation maps on the STS are readily differentiated at a coarse granular level, despite a strict topography having not yet been identified. Further work is required to identify what features have greatest potential to improve the utility of atlases in replacing functional localizers.