Computed tomography (CT) enables precise esophageal cancer diagnosis through high-resolution lesion segmentation, supporting personalized treatment. Accurate segmentation supports precise diagnosis and personalized treatment. However, existing methods often fail to capture both small and large cancerous lesions simultaneously due to their complex morphology and size variations, resulting in suboptimal performance. To address this, we propose a multi-scale global optimization network with three key components: 1) An anatomical landmark masking module (ALMM) that localizes the esophageal region using anatomical priors;2) A multi-scale dynamic fusion module (MDFM) that integrates multi-scale dilated convolutions, dynamic convolution, and dual attention for robust feature extraction;3) A global attention fusion module (GAFM) that enhances feature discrimination via channel-spatial attention and channel shuffling. We establish the private esophageal cancer CT dataset to assess the segmentation performance of our method. Experiments show our network achieves state-of-the-art performance, with a 69.75% DSC score and 19.23px HD for esophageal cancerous lesions.