Multi-object spring level sets (MUSCLE)

Abstract

A new data structure is presented for geometrically modeling multi-objects. The model can exhibit elastic and fluid-like behavior to enable interpretability between tasks that require both deformable registration and active contour segmentation. The data structure consists of a label mask, distance field, and springls (a constellation of disconnected triangles). The representation has sub-voxel precision, is parametric, re-meshes, tracks point correspondences, and guarantees no self-intersections, air-gaps, or overlaps between adjacent structures. In this work, we show how to apply existing registration algorithms and active contour segmentation to the data structure; and as a demonstration, the data structure is used to segment cortical and subcortical structures (74 total) in the human brain.