Coronary artery segmentation is a critical step in the clinical diagnosis of coronary heart disease. To tackle segmentation breaks and false segmentation caused by the thin and complex coronary vessels as well as the severe foreground-background imbalance in computed tomography angiography images, this paper proposes MSCNet, a coronary artery segmentation network with multi-scale cascade encoding and dynamic spatial context enhancement. The network constructed a multi-scale cascaded encoder using Swin Transformer and large-kernel convolutions. It sequentially modeled and fused multi-scale features by capturing long-range dependencies and local details, and reparameterized large-kernel convolutions via a spatial frequency matrix to strengthen fine detail capture. Meanwhile, a spatial transformer module was designed to dynamically guide multi-head attention learning and optimize decoding performance. On the ImageCAS dataset, MSCNet achieved an average Dice coefficient of 81.24%, which was 3.57%, 3.78%, and 3.85% higher than 3D UX-Net, SwinUNETR, and SegMamba, respectively. MSCNet effectively improves the accuracy of coronary artery segmentation and provides support for clinical evaluation.