Hierarchical 3D fully convolutional networks for multi-organ segmentation

• and Kensaku Mori  |  Hirohisa Oda1  |  Holger R. Roth  |  Kazunari Misawa  |  Masahiro Oda  |  Michitaka Fujiwara  |  Natsuki Shimizu  |  Yuichiro Hayashi

• Japan  |  Nagoya University

The model demonstrates how hierarchical deployment of 3D CNN based on a fully convolutional architecture (3D U-Net) can produce competitive results for multi-organ segmentation on a clinical CT dataset while being efficiently deployed on a single GPU. 3D FCN has been trained on manually labeled data of seven abdominal structures and has been tested for generalizability and robustness to differences in image quality and populations. It achieves high performance especially in smaller, thinner organs, such as arteries, veins and pancreas. No post-processing has been applied to FCN outputs.

Input variables : CT image
Output Variables : CT image with labeled organs

Visit Model : github.com

Additional links : arxiv.org