Semantic Segmentation of Fetal Brain Ultrasound Images Based on a General Fully Convolutional Discriminator

The fetal central nervous system (CNS) is one of the most common fetal congenital diseases in the world. It is of great significance to use deep learning methods to provide doctors with reliable auxiliary diagnosis methods. The area outside the skull halo in fetal brain ultrasound images contains a lot of irrelevant information and has fuzzy boundaries, which is not conducive to the classification or recognition tasks of ultrasound images. This paper aims to study the effect of generative adversarial networks based on universal fully convolutional discriminators on the automatic segmentation results of skull halo in fetal brain ultrasound images. This paper proposes a method for automatic segmentation of skull halo in fetal brain ultrasound images using a generative adversarial network based on a universal fully convolutional discriminator. Inspired by the concept of generative adversarial networks, a new semantic segmentation network based on a universal discriminator is constructed. In order to verify the universality of the discriminator, the semantic segmentation network uses both Deeplabᵥ2 and Attention U-net networks as generators to generate probability maps of segmentation results. A universal fully convolutional discriminator is also designed to let it learn to distinguish whether the probability map input to the discriminator network is real data or segmentation results. Experimental results on the dataset of automatic measurement of fetal head circumference in ultrasound imaging demonstrate the effectiveness of the algorithm. Compared with the baseline of Deeplabᵥ2, the segmentation accuracy is significantly improved, and the Attention U-net also has a similar improvement. Generative adversarial network based on universal fully convolutional discriminator can effectively improve the accuracy of automatic segmentation of skull halo in fetal brain ultrasound images.