Medical image segmentation plays a crucial role in the field of medical image processing, assisting doctors in detecting and evaluating lesions. Owing to the irregular shape of the tumor, uneven internal echoes and poor edge recognition, the traditional network model will segment the tumor region incorrectly, so we are committed to optimizing the model to improve the segmentation effect. In this study, we present a novel dual-branch architecture for ultrasound image segmentation, integrating ConvMixer modules to enhance global feature modelling. The encoder employs ConvMixer blocks to capture long-range dependencies efficiently. A parallel pretrained ResNet branch generates tumor probability heatmaps, which improve the localization of subtle tumors through feature-space cross-consistency loss. Additionally, we introduce a Kullback‒Leibler (KL) divergence constraint between the primary and auxiliary decoders to achieve deep feature alignment while preserving structural heterogeneity. This approach significantly improves the robustness of the model in complex lesion scenarios and enhances its clinical reliability.