Marie Curie Action

Algorithms

Algorithm for Fast and Automatic Liver Segmentation

developed at IFC-CNR Lecce, Italy (Sergio Casciaro and Laurent Massoptier)

Image segmentation consists of partitioning an image into a number of regions where their points share specific attributes (same image intensity for example). For medical applications, these structures can be anatomical or pathological. In the case of a malignant tumor inside the liver, three different treatments are possible (resection, cryo-therapy, radio-frequency ablation). However, the same information is needed from the point of view of the segmentation. The liver surface needs to be carefully delineated, the lesions and the vessel tree inside the liver are required to be precisely localized and segmented.

Figure 1

Automatic image segmentation is a challenging task because of images variability and also because algorithms should decide for each pixel if it belongs or not to the targeted structures. Concerning the liver, it is the biggest organ in the abdomen with a firm consistency, but it is easily deformed by other organs. Therefore its shape easily varies from one patient to another. Moreover, it shares the same image intensity values with other nearby organs. So, their boundaries are not always clear and sharp on medical images.

Figure 2

Fortunately, various image processing techniques are a possible candidate to rapidly and automatically segment the liver and its internal structures. They can be regrouped into two main families: the region approach looking for common properties between regions, and the contour approach looking for discontinuities or intensity differences. We choose to combine several techniques in a three-dimension coarse-to-fine approach. Firstly, the algorithm looks for the liver surface: gradient vector field active contours are used to smooth and refine an initial segmentation resulting from an adaptive thresholding step (Figures 1). Then, the inside of the liver is classified by clustering into three distinct objects: parenchyma, vessels and lesions (Figure 2). The solution we implemented is fully automatic and gives good results with MRI or CT image volumes indifferently. With a small processing time, pre-operative and treatment quality control phases could be the context of use of this algorithm. For the intra-operative phase, these pre-operative segmented data will be registered to real-time ultrasound acquisitions.

Contacts