Cascaded Fully Convolutional DenseNet for Automatic Kidney Segmentation in Ultrasound Images

Although the extraction of traditional features has achieved good performance in kidney segmentation, the automatic kidney segmentation in whole ultrasound images is still challenging. Due to the characteristics of the kidney ultrasound images such as obvious noise, severe redundancy information and low image contrast, it is difficult to achieve accurate segmentation of the kidney by directly using the fully convolutional dense network (FC-DenseNet). Therefore, we propose a cascaded FC-DenseNet, that is, a model of coarse-to-fine segmentation. By training the coarse segmentation model, prediction probability that the pixel becomes the correct class label is calculated, and the proposal region of the kidney in the ultrasound image is obtained. By training the fine segmentation model, the proposal region of kidney is finely segmented and restored back to the original size. Compared with the original FC-DenseNet, cascaded FC-DenseNet has better performance in kidney segmentation. The mean intersection over union is increased by 1.43%, and the samples that are difficult to segment could be better segmented.