A New Adaptive Variational Model for Liver Segmentation with Region Appearance Propagation

Liver segmentation is a crucial step for aiding in liver surgery. Due to intensity overlapping, blurred edges, large variability in shape and appearance, and complex context with clutter features, it is still a challenging task. In this paper, we address this problem with an integrated variational model based on the idea of adaptive region growing and region appearance propagation, with which we can focus on the target liver region regardless of the complex but uninterested backgrounds. Our model consists of an edge based term and two novel region based terms, with which both region intensity and appearance information are integrated and weak liver boundaries can be stably delineated. An adaptive weight is introduced to spatially balance them and control their respective advantages and disadvantages. Moreover, the proposed model is robust to parameters, initialization and noises, and can greatly alleviate the requirement of the scanning protocol and data quality. Last but not least, our model is a nearly automatic one which needs only an arbitrary initialization inside the liver. While segmenting slice by slice neglects the consecutiveness between slices, we directly segment the liver from the 3D volume data. Experimental results show that the liver can be accurately and effectively distinguished, and vessels can also be simultaneously isolated with accuracy. Our system is promising for stable practical use and can be also used to segment other abdominal organs.