Multi-scale structured CNN with label consistency for brain MR image segmentation

Abstract

In this paper, a novel method for brain MR image segmentation has been proposed, with deep learning techniques to obtain preliminary labelling and graphical models to produce the final result. A specific architecture, namely multi-scale structured convolutional neural networks (MS-CNN), is designed to capture discriminative features for each sub-cortical structure and to generate a label probability map for the target image. Due to complex background in brain images and the lack of spatial constraints among testing samples, the initial result obtained with MS-CNN is not smooth. To deal with this problem, dynamic random walker with decayed region of interest is then proposed to enforce label consistency. Comprehensive evaluations have been carried out on two publicly available data-sets and experimental results indicate that the proposed method can obtain better segmentation quality efficiently.

Keywords:

• Deep learning
• convolutional neural networks
• label consistency
• brain MR image

Notes

No potential conflict of interest was reported by the authors.

1 http://www.nitrc.org/projects/ibsr

2 http://www.loni.usc.edu/atlases/Atlas_Detail.php?atlas_id=12

Additional information

Funding

This work was supported in part by the K. S. Lo Foundation and the Research Grants Council of Hong Kong General Research Fund [grant number 16203115].