Fast Watersnakes: an improved image segmentation framework

Abstract

Fissure detection is an important task in the interpretation and diagnosis of pathologies present in lung CT images and has a lot of challenges in terms of speed and accuracy. In this paper, a new method called Fast Watersnakes to detect and segment the fissures has been proposed. Fast Watersnakes integrates the speed of Fast Watershed and the smoothness of active contours to obtain the desirable segmentation. Fast Watershed based on chain codes provides a prominent solution to the over-segmentation problem of morphological watersheds. However, there is no control of smoothness in the segmentation results. Existing methods use watershed line and contour length to incorporate smoothness, but there are no watershed lines in Fast Watersheds. The proposed method addresses Fast Watersheds as energy minimisation function. Experimental results show that the proposed method overcomes the over-segmentation problem and shows a considerable reduction in root mean square (RMS) error values when tested with lung CT images. The proposed method gives an RMS error range of 1·98±1·60 mm for fissure segmentation when compared with expert observations.

Keywords:

• Lung CT images
• Fissures
• Segmentation
• Watershed
• Fast Watersnakes

Acknowledgements

The authors would like to thank Dr P. Senthil Kumar, Consultant Radiologist, Saravana Scans, Salem, India, for providing valuable suggestions and results of manual tracing, and Dr. R. Saravanan, Director, Deepam Hospitals, Chennai, India, for the data set used in this work.