<p>Ultrasound imaging is widely used for diagnosing various medical conditions. However, lesion segmentation in ultrasound images is challenging due to low contrast, noise, blurred boundaries, and variability in lesion characteristics. To address these issues, we propose a Global Response Double U-shaped Network, a hybrid CNN-Transformer architecture designed for lesion segmentation in ultrasound images. The encoder-decoder backbone is constructed using a U-shaped Dilated Convolution module, which effectively captures fine-grained local features and enhances boundary delineation, particularly under low-contrast conditions. To improve global context modeling, this paper proposes the Global Response Transformer Block in the bottleneck, enabling the network to capture long-range dependencies and structural variability in lesion appearance. By modeling interactions across distant regions, the block more effectively captures the variability in lesion shape, size, and location, enhancing segmentation accuracy for complex and irregular structures in ultrasound images. Furthermore, we design a Multi-Scale Linear Attention Gate to refine skip connections by emphasizing salient features and suppressing redundancy, thereby mitigating noise interference and improving decoding efficiency. By suppressing speckle noise and enhancing critical features, this mechanism improves segmentation accuracy and ensures robustness in complex ultrasound imaging scenarios. The proposed method has been extensively evaluated on publicly available ultrasound image datasets, including breast and thyroid lesion data, demonstrating its effectiveness and robustness in segmenting complex and low-contrast lesions under real-world imaging conditions.</p>