Enhanced Aortic CT Synthesis Based on Multiscale Information Fusion

For aortic dissection, we aim to synthesize contrast-enhanced computed tomography (CE-CT) from noncontrast-enhanced CT (NCE-CT), avoiding the possible side effects caused by contrast agents in the acquisition of CE-CT. We propose a novel generative adversarial network (GAN) based on multiscale information fusion, named as MIF-GAN. The generator adopts a dual-way Unet architecture, leveraging dense connections to effectively fuse adjacent blocks. Additionally, it incorporates attention-guided feature selection, optimizing the combination of high-resolution and shallow features. Furthermore, a multikernel residual convolution module enhances multiscale contextual features, boosting the overall performance of the model. Experimental results demonstrate that this method can synthesize CE-CT images of high quality and similarity for aortic dissection. With this approach, we hope to reduce the cost for CE-CT scans, thereby reducing patient discomfort and healthcare costs.