Abstract
Introduction. Delineation of retinotopic map boundaries in human visual cortex is a time-consuming task. Automated methods based on anatomy (cortical folding pattern; Benson et al., 2014; DOI:10.1016/j.cub.2012.09.014) or a combination of anatomy and retinotopic mapping measurements (Benson & Winawer, 2018; DOI:10.7554/eLife.40224) exist, but human experts are more accurate than these methods (Benson et al., 2021; DOI:10.1101/2020.12.30.424856). Convolutional Neural Networks (CNNs) are powerful tools for image processing, and recent work has shown they can predict polar angle and eccentricity maps in individual subjects based on anatomy (Ribiero et al., 2021; DOI:10.1016/j.neuroimage.2021.118624). We hypothesize that a CNN could predict V1, V2, and V3 boundaries in individual subjects with greater accuracy than existing methods. Methods. We used the expert-drawn V1-V3 boundaries from Benson et al. (2021) of the subjects in the Human Connectome Project 7 Tesla Retinotopy Dataset (Benson et al., 2018; DOI:10.1167/18.13.23) as training (N=135) and test data (N=32). We constructed a U-Net CNN with a ResNet-18 backbone and trained it with either anatomical (curvature, thickness, surface area, and sulcal depth) or functional (retinotopic) maps as input. Results. CNN predictions out-performed other methods. The median dice coefficients between predicted and expert-drawn labels from the test dataset for the CNNs trained using anatomical and functional data were 0.77 and 0.90, respectively. In comparison, coefficients for existing methods based on anatomical or anatomical plus functional data were 0.70 and 0.72, respectively. These results demonstrate that even with a small training dataset, CNNs excel at accurately labeling visual areas on human brains in an automated fashion. This method can facilitate vision science neuroimaging experiments by making an otherwise difficult and subjective process fast, precise and reliable.