Dual-domain latent alignment for MRI-to-CT synthesis via target-aware manifold learning.

Computed tomography (CT) is essential for clinical diagnosis and radiotherapy planning but involves ionizing radiation, motivating the synthesis of CT images from magnetic resonance imaging (MRI). Most existing MRI-to-CT synthesis methods rely on fully paired supervision, which is difficult to scale in practice and restricts the availability of diverse training data. While unpaired CT samples can provide additional anatomical priors, current approaches typically exploit them as generic style references, lacking explicit mechanisms to preserve anatomical geometry. In this work, we propose DELA-Net, a Dual-Encoder Latent Alignment Network that reformulates MRI-to-CT synthesis as a target-aware latent manifold alignment problem. DELA-Net introduces a geometric reference selection strategy to identify informative unpaired CT samples as latent anchors, enabling structured alignment between paired and unpaired domains. The optimization is decomposed into two coupled alignment tasks - MRI-to-unpaired CT and paired-to-unpaired CT - facilitated by a latent space modulator that adapts MRI representations to the CT manifold while maintaining structural fidelity. Experiments on three benchmark datasets provide promising evidence that DELA-Net achieves a favorable balance across synthesis quality, cross-modality task translation, and bone-related structural preservation in downstream task-driven evaluation. The source code is publicly available at: https://github.com/kennysyp/DELA-Net. The GitHub repository is now publicly accessible.