Convolutional networks for kidney segmentation in contrast-enhanced CT scans

Abstract

Organ segmentation in medical imagery can be used to guide patient diagnosis, treatment and follow ups. In this paper, we present a fully automatic framework for kidney segmentation with convolutional networks (ConvNets) in contrast-enhanced computerised tomography (CT) scans. In our approach, a ConvNet is trained using a patch-wise approach to predict the class membership of the central voxel in 2D patches. The segmentation of the kidneys is then produced by densely running the ConvNet over each slice of a CT scan. Efficient predictions can be achieved by transforming fully connected layers into convolutional operations and by fragmenting the maxpooling layers to segment a whole CT scan volume in a few seconds. We report the segmentation performance of our framework on a highly variable data-set of 79 cases using a variety of evaluation metrics.

Keywords:

• Medical imaging and visualisation
• image processing and analysis

Acknowledgements

The authors would like to thank the developers of Theano (Bergstra et al. 2010; Bastien et al. 2012).

Notes

No potential conflict of interest was reported by the authors.

1 Note that the output size is actually because no padding was used before applying the convolutional operations.

Additional information

Funding

This paper was supported in part by the Canada Research Chair in Medical Imaging and Assisted Interventions, the Natural Sciences and Engineering Research Council of Canada, the MEDITIS training program and the NSERC Engage program.