Intravascular ultrasound (IVUS) is the preferred modality for capturing real-time and high resolution cross-sectional images of the coronary arteries, and evaluating the stenosis. Accurate and real-time segmentation of IVUS images involves the delineation of lumen and external elastic membrane borders. In this paper, we propose a two-stream framework for efficient segmentation of 60 MHz high resolution IVUS images. It combines shallow and deep networks, namely, CSDN. The shallow network with thick channels focuses to extract low-level details. The deep network with thin channels takes charge of learning high-level semantics. Treating the above information separately enables learning a model to achieve high accuracy and high efficiency for accurate real-time segmentation. To further improve the segmentation performance, mutual guided fusion module is used to enhance and fuse both different types of feature representation. The experimental results show that our CSDN accomplishes a good trade-off between analysis speed and segmentation accuracy.