Shape-Adaptive Guidance Signal for Interactive Cortical Sulcal Labeling

ICLR 2026 Conference Submission18472 Authors

Keywords: Spherical CNN, Interactive Segmentation, Cortical Sulci, Lateral Prefronal Cortex

Abstract: Image segmentation is a fundamental task in image data analysis that assigns a semantic label to enhance the understanding of imaging data. In the context of neuroimaging data, the accurate labeling of cortical sulci is crucial for providing deep insights into the link between cortical folding patterns and cognitive functions. Yet, fully automatic methods often struggle with labeling small and shallow sulci due to their high anatomical variability and the scarcity of annotated training data. In this context, interactive segmentation may offer a promising alternative by incorporating minimal human input to refine labels. However, recent learning-based interactive approaches often rely on 2D projections of surface data, typically designed for generic and relatively small 3D meshes. This dimensional simplification inherently limits their ability to capture subtle folds and deeply buried structures of cortical surfaces. In this paper, we introduce a shape-adaptive guidance signal for interactive cortical sulcal labeling using spherical convolutional neural networks. Thanks to the use of spherical mapping, our approach preserves structural information without the need for sacrificing anatomical details. To effectively encode user clicks along cortical folding patterns, we solve the eikonal equation with a speed function that incorporates the mean curvature of the cortical surface unlike conventional encoding schemes using equidistance. This curvature-aware signal captures fine-grained anatomical details to guide the neural network focus on the intended refinement. Experimental results on 72 subjects with 17 sulci on the lateral prefrontal cortex show that even a single click using the proposed encoding scheme outperforms fully automatic methods and equidistance schemes, while achieving both efficiency and improved labeling accuracy.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 18472