Developing accurate boundary extraction methods for ultrasound images of the kidney is a significant challenge on account of the lack of boundaries or blurred boundaries. This study puts forward a method from rough to fine from four aspects of innovation. First, we use the characteristics of a principal curve (PC) to adjust the generated boundary outline automatically as well as utilize a neural net to cut down the errors generated by the model. Second, an initial boundary extraction phase is performed using a deep parallel fusion training network, with a parallel structure to boost the segmentation achievement of deep learning. Third, we designed an automatic system for extracting the polygonal boundary, which used a mean-shift cluster-based method to substitute for the traditional PC-based methodologies consisting of the projection and vertex optimization procedure. Fourth, an explicable mathematic mapping function of the renal outline was developed, which is expressed via parameters of the neural network. Finally, we carried out several experiments to demonstrate the performance of our approach.