Segment anything model guided dual-mask framework for anatomically faithful medical image translation

Image-to-image translation enables cross-modal synthesis in medical imaging, such as MRI-to-CT conversion. While generative adversarial networks (GANs) have demonstrated strong capabilities in these tasks, existing methods often struggle to preserve anatomical accuracy, particularly in complex structures, due to the absence of explicit anatomical constraints. To overcome this limitation, we propose a GAN-based medical image translation framework that incorporates segmentation priors from the Segment Anything Model (SAM). Our framework has two components: (1) a SAM-guided multi-resolution feature module and (2) a dual-mask refinement mechanism. Beyond simple foreground–background separation, this mechanism encourages hierarchical, specialized representations. It explicitly enforces alignment by using the overall anatomical silhouette and selectively preserves internal fine textures and edges. Our SAM-guided approach improves spatial coherence and anatomical fidelity in the translated images. Experimental evaluations on MRI-to-CT cross-modal translation tasks confirm that our proposed method substantially outperforms existing baseline approaches in terms of both anatomical accuracy and overall image quality.