Abstract:CT synthesis based on ultra-short echo-time magnetic resonance imaging (UTE MRI) can eliminate the risk of ionizing radiation and has important clinical and ethical value for radiation-sensitive populations, such as children, pregnant women, and patients requiring long-term follow-up. To address the common challenges of anatomical misalignment and texture loss in existing cross-modal synthesis methods, this study proposes CASA-Net. The model incorporates a structure-texture aware multi-scale cross-axis attention (ST-MCA) module in the generator to enhance global structural modeling and multi-scale contextual modeling capabilities. Meanwhile, an SAPatchGAN discriminator based on spectral normalization and self-attention is designed to improve adversarial supervision performance for texture and structural consistency. Validation using 100 clinical cases shows that CASA-Net achieves SSIM, PSNR, and MAE values of 0.870±0.009, 23.962±0.767 dB, and 5.796±0.778 HU, respectively, outperforming mainstream methods such as CycleGAN and ResViT. The proposed method can effectively synthesize high-fidelity CT images, providing key technical support for radiation-free follow-up in patients with chronic lung diseases.