Learned snakes for 3D image segmentation

Lihong Guo, Yueyun Liu, Yu Wang, Yuping Duan*, Xue-Cheng TAI

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

1 Citation (Scopus)

Abstract

Snakes or active contour models are classical methods for boundary detection and segmentation, which deform an initial contour (for 2D image) or a surface (for 3D image) towards the boundary of the desired object. Such snakes models are ideal choices for handling medical image segmentation problems since they are very efficient and require fewer memories by solely tracking the explicit curves or surfaces. However, traditional snakes models solved by the level set method suffer from numerical instabilities and are usually difficult to deal with topological changes. In this paper, we propose a learned snakes model for 3D medical image segmentation, where both the initial and final surfaces are estimated using deep neural networks in end-to-end regimes. The merit of our learned snakes model is that we can realize 3D segmentation by finding a 2D surface based on 2D convolutional neural networks rather than using 3D networks or cutting the volume into 2D slices. Experiments on the Medical Segmentation Decathlon spleen dataset against both 2D- and 3D-based networks demonstrate our model achieving the state-of-the-art accuracy and efficiency, which not only enjoys a 1% higher DSC but also saves more than 90% computational time compared to the well-established elastic boundary projection model Ni et al. [1].

Original languageEnglish
Article number108013
JournalSignal Processing
Volume183
DOIs
Publication statusPublished - Jun 2021

Scopus Subject Areas

  • Control and Systems Engineering
  • Software
  • Signal Processing
  • Computer Vision and Pattern Recognition
  • Electrical and Electronic Engineering

User-Defined Keywords

  • 3D segmentation
  • Active contour
  • Convolutional neural network
  • Snakes
  • Surface evolution

Fingerprint

Dive into the research topics of 'Learned snakes for 3D image segmentation'. Together they form a unique fingerprint.

Cite this