associated_data_Investigating the Effectiveness of clDice Loss for Road Crack Segmentation

Abstract

Accurate segmentation of tubular, network-like structures, such as vessels,\nneurons, or roads, is relevant to many fields of research. For such structures,\nthe topology is their most important characteristic; particularly preserving\nconnectedness: in the case of vascular networks, missing a connected vessel\nentirely alters the blood-flow dynamics. We introduce a novel similarity\nmeasure termed centerlineDice (short clDice), which is calculated on the\nintersection of the segmentation masks and their (morphological) skeleta. We\ntheoretically prove that clDice guarantees topology preservation up to homotopy\nequivalence for binary 2D and 3D segmentation. Extending this, we propose a\ncomputationally efficient, differentiable loss function (soft-clDice) for\ntraining arbitrary neural segmentation networks. We benchmark the soft-clDice\nloss on five public datasets, including vessels, roads and neurons (2D and 3D).\nTraining on soft-clDice leads to segmentation with more accurate connectivity\ninformation, higher graph similarity, and better volumetric scores.\n

Ähnliche Arbeiten