Automatic liver segmentation on CT images combining region-based techniques and convolutional features

Precise automatic liver segmentation plays an important role in computer-aided diagnosis of liver pathology. Despite many years of research, this is still a challenging task, especially when processing heterogeneous volumetric data from different sources. This study focuses on automatic liver segmentation on CT volumes proposing a fusion approach of traditional methods and neural network prediction masks. First, a region growing based method is proposed, which also applies active contour and thresholding based probability density function. Then the obtained binary mask is combined with the results of the 3D U-Net neural network improved by GrowCut approach. Extensive quantitative evaluation is carried out on three different CT datasets, representing varying image characteristics. The proposed fusion method compensates for the drawbacks of the traditional and U-Net based approach, performs uniformly stable for heterogeneous CT data and its performance is comparable to the state-of-the-art, therefore it provides a promising segmentation alternative.