Liver segmentation is an important prerequisite for surgery planning. The manual segmentation of the liver is very time consuming, so it is desired to develop a method that can precisely segment the liver without any human interaction. In spite of several decades of research and many key advances, no fully automatic and robust so- lutions are commercially available. This is due to the fact that some issues still have to be addressed: high shape variation due to natural anatomical variation, disease, or pre- vious surgical interventions; inhomogeneous grey value appearance caused by tumors or metastasis; low contrast to neighbouring structures/organs like stomach, heart, muscle.
In this thesis a fully automatic method for healthy liver segmentation on contrast-enhanced CT images is proposed . The method combines segmentation based on a SSM (Statistical Shape Model) and a free form segmentation step based on optimal graph searching to take advantage of the potentials of each single technique. By the use of a SSM, prior knowledge about the typical liver shape is incorporated into the segmentation process to constrain it where the image information is not reliable. Any segmentation, resulting from SSM adaptation is an instance of the shape model. For this reason, new and unknown shapes cannot be captured by the SSM. Due to the high liver variability in shape and size a SSM segmentation is not sufﬁcient to get good result. To overcome this limitation a free form deformation step is applied to the SSM segmentation result.
The used appearance model is based on proﬁles running perpendicular to the surface et each vertex. A cost function based method estimates the position of liver boundary within each proﬁle. A cost is computed for each sample of all intensity proﬁles. During SSM segmentation step the position of the minimum yields the displacement vector for each vertex. In the free form deformation step the liver boundary is obtained via optima graph cut global minimization, while respecting constraints on shape preservation. The cost assigned to each intensity proﬁle sample depends on the matching degree of the sample and its neighbours with certain features of the liver boundary. These features are collected in a ﬂexible model. The model describes the liver boundary in terms of shape, size, intensity and gradients. The model parameters are set analysing the intensity proﬁles of reference surfaces manually segmented by radiological experts.
The method is evaluated on 87 test scans which do not present evident lesions, by comparing the automatically detected liver volumes to the liver boundaries manually traced by radiology experts. The algorithm takes 9 minutes per liver on an Intel 2.66 GHz processor. Despite some outliers due to a bad initialization of the SSM pose in the current volume, the method shows an high accuracy level. Precisely, the ﬁve metrics used to evaluate the method that are OE (Volumetric Overlap Error), AVD (Absolute Vol- ume Difference), AD (Average Symmetric Surface Distance), RMS (Root Mean Squared Symmetric Surface Distance), MD (Maximum Symmetric Surface Distance) present re- spectively a median of 7.94%,2.14%, 0.12 mm, 0.26 mm and 2.30 mm. Inferior values iiiof these metrics in totally automatic liver segmentation have not been found in any other approach present in literature.
According to the results, the approach presented in this work represents a promising step towards the automatic liver segmentation from CT images. The segmen- tation method is characterized by high reliability and accuracy being superior to other methods previously reported in literature. Thanks to an optimized implementation, the low computational load is compatible with the clinical work-ﬂow in the surgical planning stage. While addressing part of the listed issues, further analysis appears however to be mandatory before the method is included into a clinical tool.