CT-to-MRI synthesis without 3D diffusion: StaX-Net shares interleaved 2D blocks and reaches an abdominal perceptual distance of 642.1, below Pix2Pix
Synopsis
Cross-modality medical image translation has been troubled by anatomically inconsistent synthesis, and standard generative adversarial networks update generator and discriminator together, letting opposing gradients pull on shared representations. StaX-Net interleaves multi-scale computational blocks between the generator and discriminator, lets the discriminator update first while shared blocks are frozen during the generator step, and uses target MRI only during training to guide a feature transform. On the Abdomen/Pelvis, Head and Neck, and Thorax subsets of SynthRAD 2025 it is compared with five baselines, reaching an Abdomen/Pelvis VGG perceptual distance of 642.1, below Pix2Pix at 991.2 and UNet-AE at 1687.9.
Interpretation
CT-to-MRI synthesis can now be done inside a two-dimensional generative adversarial network rather than relying on an explicit 3D diffusion architecture. Earlier coupled generative adversarial networks updated shared parameters from both generator and discriminator at once, so two opposing objectives pulled the shared representation in different directions. On the Abdomen/Pelvis, Head and Neck, and Thorax subsets of SynthRAD 2025, StaX-Net is compared with Pix2Pix, Vanilla cGAN, UNet-AE, CycleGAN, and an adapted MedViT-cGAN; its Abdomen/Pelvis VGG perceptual distance is 642.1, below Pix2Pix at 991.2 and UNet-AE at 1687.9.
Shared blocks appear interleaved in both pathways, so generator and discriminator use the same parameterized block at corresponding stages while keeping separate activations and dedicated layers. Sharing parameters does not mean the two pathways process an identical feature tensor, nor does it collapse into one shared encoder followed by independent branches. The discriminator acts as leader and updates the shared blocks and its own parameters first; when the generator acts as follower the shared blocks are frozen and only dedicated synthesis parameters update, with asymmetric learning rates compensating for the generator's restricted update space.
During training the target MRI enters only through a dedicated projection into an auxiliary loss, distilling target-domain structure into a feature transform that depends on CT alone. The target image is never concatenated into the forward input and is entirely absent at inference, where the FAST layer performs a deterministic source-only transform. That auxiliary loss compares projected target features with the dynamic prior bank in feature space rather than raw pixel space, and after training the inference stage needs no target-dependent adaptation.
Perspective
Evaluation is confined to the Abdomen/Pelvis, Head and Neck, and Thorax subsets of the SynthRAD 2025 Grand Challenge dataset, split at the subject level into 80% training, 10% validation, and 10% testing, all paired CT and MRI volumes. Readers should treat it as evidence that paired cross-modality translation is feasible in a 2D architecture, not as a clinical deployment conclusion.
The authors explicitly do not claim convergence to a strict Stackelberg equilibrium, which would require computationally prohibitive hypergradients, and describe the asymmetric update as a stabilizing heuristic. On the Thorax subset StaX-Net reaches 2D PSNR 18.67 and 3D PSNR 18.37, below UNet-AE at 19.26 and 18.90; the authors' discussion of deterministic architectures favoring smooth outputs addresses UNet-AE's Head and Neck PSNR and Abdomen/Pelvis MAE, not specifically this Thorax gap.