An Efficiency Coarse-to-Fine Segmentation Framework for Abdominal Organs Segmentation

Cancan Chen, Weixin Xu, Rong guo Zhang

Abstract: U-Net has been proved as the most successful segmentation architecture for medical image processing in recent years. Based on this, ResUNet imported ResBlock with skip connection focuses more on the contextual information. In this work, we adopt the 3D ResUNet to build a whole-volume-based coarse-to-fine segmentation framework for the abdominal multi-organs segmentation task, and the mean Dice Similarity Coefficient (DSC) of the segmentation results has achieved 87.67\%, the mean Normalized Surface Dice (NSD) has achieved 93.16\% on the FLARE2022 validation set. Besides, for each case on the FLARE2022 validation set, the average running time is 19.5614 seconds, and the max gpu memory consumption is 2657 MB