Schedule of Events | Symposia

Hemispheric biases in automatic atlas-based cortical parcellations exaggerate surface area lateralization

Poster Session D - Monday, March 31, 2025, 8:00 – 10:00 am EDT, Back Bay Ballroom/Republic Ballroom

Yinuo Liu¹, Ja Young Choi², Tyler K. Perrachione¹; ¹Boston University, ²Northwestern University

Anatomical asymmetry is a hallmark of the human brain and may reflect hemispheric differences in its functional organization. Powerful software like FreeSurfer can automate neuroanatomical measurements and facilitate studies of hemispheric asymmetry. However, surface area asymmetry results from FreeSurfer’s default processing pipeline are curiously consistent across diverse samples. Here, we demonstrate that systematic biases in these measurements exist when using the default processing pipeline in FreeSurfer. We compared surface area asymmetry measured from reconstructions of natural brains vs. reconstructions of the same scans after flipping their left-right orientation. The default pipeline returned implausible asymmetry patterns across original and flipped brains (i.e., structures that were always left- or right-lateralized). In contrast, manual labelling of key structures yielded expected reversals of left/right lateralization in flipped brains. A linear SVM model trained to classify left vs. right hemispheres based on automatic regional surface area measurements was suspiciously accurate (>98%) for natural brains but mislabeled the vast majority (>80%) of flipped brains, further indicating biases in these measurements that do not reflect the underlying neuroanatomy. Notably, these biases are greatest in key speech and language regions. We determined that these biases result from discrepancies in how regional labels are defined in the default hemisphere-specific atlases. We further demonstrate how these biases can be ameliorated by using the symmetric registration templates and parcellation atlases available from FreeSurfer, but separate from the default pipeline. These results underscore the need for validating bias-free neuroanatomical measurements, particularly when studying regions likely to exhibit hemispheric lateralization.

Topic Area: METHODS: Neuroimaging