Hierarchical 3D fully convolutional networks for multi-organ segmentation

• DOI: 10.48550/arxiv.1704.06382
• Publication Date: 2017-01-01
• Authors: Holger R. Roth; Hirohisa Oda; Yuichiro Hayashi; Masahiro Oda; Natsuki Shimizu; Michitaka Fujiwara; Kazunari Misawa; Kensaku Mori
• Keywords: Dice

Recent advances in 3D fully convolutional networks (FCN) have made it feasible to produce dense voxel-wise predictions of full volumetric images. In this work, we show that a multi-class FCN trained on manually labeled CT scans seven abdominal structures (artery, vein, liver, spleen, stomach, gallbladder, and pancreas) can achieve competitive segmentation results, while avoiding the need for handcrafting features or training organ-specific models. To end, propose two-stage, coarse-to-fine approach trains an model roughly delineate organs interest first stage (seeing $\sim$40% voxels within simple, automatically generated binary mask patient's body). We then use these first-stage define candidate region will be used train second FCN. This step reduces number has classify $\sim$10% maintaining recall high $>$99%. second-stage now focus more detailed organs. respectively utilize validation sets consisting 281 50 clinical Our hierarchical provides improved Dice score 7.5 percentage points per organ average our set. furthermore test models completely unseen data collection acquired at different hospital includes 150 with three anatomical labels (liver, pancreas). such challenging as pancreas, improves mean from 68.5 82.2%, achieving highest reported dataset.