A Two-Stage Framework for Kidney Segmentation in Ultrasound Images.

In this paper, we propose an algorithm framework for kidney segmentation in ultrasound images. Due to the characteristics of kidney ultrasound images, such as high noise, heterogeneous structure, low contrast, multiple artifacts, and relatively fixed shape, accurately segmenting clear and complete kidney structures from the images is still a challenging task. Our framework consists of two parts: shape aware dual-task multi-scale fusion network and self-correction. The first part uses a U-shape structure with multi-scale feature cross fusion skip connections to perform segmentation and can simultaneously predict the segmentation map and the shape-aware prediction level set function, we use a dual-task consistency module to constrain the shape of the segmentation map, enabling the network to learn the target area more accurately, through dual-task consistency supervised learning, we obtain a pre-trained model. The second part uses an iterative aggregation strategy for the pre-trained model to optimize it and reduce the noise and other issues in the prediction results. Experimental results show that our algorithm framework outperforms several state-of-the-art methods on kidney ultrasound datasets.