Mathematical Modeling Method For Liver Segmentation In CT Images

Accurate segmentation of liver from the abdominal CT scans always is a hot topic in medical image segmentation, and also a challenging task. Focus on this task, a semi-automatic model and a fully automatic method is proposed in this paper.Low contrast, weak boundary and noise often exist in the liver CT scans.Furthermore, the presence of tumor and intensity inhomogeneity in CT scans may part the liver into two sub-regions with distinct appearances. Specially,accurate liver segmentation becomes more challenging when the tumor locates on the liver surface. An apparently simple solution may run some model twice,each segmentation process for each sub-region. However, it is time-consuming and memory consumption. Under the framework of active contour, we construct a joint model of delineating the whole liver with a single level set representation.The final result and computational efficiency is improved compared to segmenting every part independently. The level set evolution on the multiple regions is guided to the desired boundary by a novel geodesic selection scheme. As a result, in each sub-region the most relevant appearance and boundary knowledge are automatically used. Besides, a weighted histogram for the local appearance description is introduced for the precise boundary detection.The proposed geodesic selection based model need some user intervention,which are undesirable in clinic usage. Thus, in this paper, we also develop a deep learning algorithm with graph cut refinement to automatically segment the liver in CT scans. This method consists of two main steps: (i) simultaneously liver detection and probabilistic segmentation using 3D convolutional neural net-work; (ii) accuracy refinement of the initial segmentation with graph cut and the previously learned probability map. The main advantage of our method is that it does not require any user interaction for initialization. Thus, the proposed method can be performed by nonexperts. In addition, our work is one of the early attempts of employing deep learning algorithms for 3D liver segmentation.All proposed approaches were validated on public databases. And the quan-titative and qualitative results show that the two proposed methods performed well on accurate liver delineation from CT scans.