Ultrasound image is an essential basis for clinicians to diagnose and make treatment plans. Due to the limitation of imaging mechanism, ultrasound images are usually accompanied with speckles and noise, resulting in a adrop in image quality. However, image quality is a key factor for lesion region segmentation and pathological analysis. Therefore, suppressing speckles, artifacts and noise is of great significance for diagnosing. In this paper, we propose a blind super-resolution guided network(BSG-Net) for improving ultrasound image segmentation. Our BSG-Net consists of a novel self-supervised cycle blind super-resolution generative adversarial network(CycleBSR-GAN) for ultrasound image reconstruction and a W-shape network(W-Net) for lesion segmentation. To avoid redundant textures and to ensure the cycle-consistency between the ground truth and the generated image, we adopt the structure of Cycle-GAN. Our CycleBSR-GAN consists of two generators and two attention based discriminators, which aims to enhance ultrasound image quality in terms of both numerical criteria and visual results. The W-Net comprises a multi-scale input layer, EfficientNetB3 encoder blocks, atrous spatial pyramid pooling(ASPP) blocks, attention module and a our proposed parallel atrous convolution(PAC) block. Because of its powerful feature extraction capability, we use the W-Net as our low to high resolution generator and also perform the segmentation task. In order to simulate the characteristics of uneven grayscale, speckles and noise in real-world ultrasound images, we use a random multiple combination degradation strategy(RMCDS) to consturct training sample pairs. The experimental results have shown that our BSG-Net paired with RMCDS achieved state-of-the-art(SOTA) performance on the TNUI-2021 datasets.