Centerline Boundary Cube Loss for Vascular Segmentation
Authors: Pengcheng Shi, Jiesi Hu, Yanwu Yang, Zilve Gao, Wei Liu, Ting Ma
Summary: Vascular segmentation in medical imaging performs an important function in analysing morphological and practical assessments. Conventional strategies, just like the centerline Cube (clDice) loss, guarantee topology preservation however falter in capturing geometric particulars, particularly underneath translation and deformation. The mixture of clDice with conventional Cube loss can result in diameter imbalance, favoring bigger vessels. Addressing these challenges, we introduce the centerline boundary Cube (cbDice) loss perform, which harmonizes topological integrity and geometric nuances, making certain constant segmentation throughout varied vessel sizes. cbDice enriches the clDice strategy by together with boundary-aware features, thereby bettering geometric element recognition. It matches the efficiency of the boundary distinction over union (B-DoU) loss by means of a mask-distance-based strategy, enhancing traslation sensitivity. Crucially, cbDice incorporates radius data from vascular skeletons, enabling uniform adaptation to vascular diameter modifications and sustaining stability in department development and fracture impacts. Moreover, we performed a theoretical evaluation of clDice variants (cl-X-Cube). We validated cbDice’s efficacy on three various vascular segmentation datasets, encompassing each 2D and 3D, and binary and multi-class segmentation. Notably, the strategy built-in with cbDice demonstrated excellent efficiency on the MICCAI 2023 TopCoW Problem dataset. Our code is made publicly obtainable at: https://github.com/PengchengShi1220/cbDice
