Atlas based liver segmentation using nonrigid registration with a B-spline transformation model

Abstract

Liver segmentation is an important step for the therapeutic decision making in liver surgery. However, manual segmentation is timeconsuming and tedious and so the need for accurate and robust automatic segmentation methods for clinical data arises. In this work an atlas in combination with nonrigid registration is used to segment the liver in actual clinical CT images. First, the atlas is built on twenty training images using nonrigid registration with a novel surface distance penalty. Next, this atlas is nonrigidly registered to ten test images. Currently, the user interaction is limited to the initialization of a rigid registration and to the definition of a region of interest for the nonrigid registration. Future work will focus on replacing the remaining user interaction with fully automatic procedures. Results are promising with an average overlap error of 10.4% and an average RMS distance of 5.0mm for the ten test images. Errors occur mainly at sites where the atlas is ill-defined such as the border between the heart and the liver.