Search papers, labs, and topics across Lattice.
This paper introduces WING, a novel generative network designed to synthesize CT volumes from MRI and CBCT images by reformulating the regression target into multiple windowed representations. By leveraging the structure-deterministic nature of CT intensities, WING employs a Gated Inception Generator for multi-window predictions and a Fuse-and-Refine Transformer for detail refinement, achieving state-of-the-art performance in cross-modality synthesis. The results indicate that WING not only improves the accuracy of CT generation but also supports multi-anatomy synthesis within a single model, enhancing treatment planning in adaptive radiotherapy without additional radiation exposure.
WING achieves state-of-the-art CT synthesis from MRI and CBCT by transforming the regression target into windowed representations, significantly improving detail and accuracy.
Generating CT volumes from MRI and CBCT can improve treatment planning in adaptive radiotherapy while avoiding additional radiation exposure. However, direct regression of CT intensities is challenged by the inherently high dynamic range and long-tailed distributions, thereby averaging out sparse yet clinically important structures. To alleviate this issue, we reformulate the regression target into multiple windowed representations, leveraging the inductive prior that CT intensities are structure-deterministic and window-separable. These windowed views exhibit smoother distributions and admit structured fusion back to the full-range CT. Building on this reformulation, we introduce WING, a WINdow-prior-based Generative network comprising: 1) a new Gated Inception Generator to produce multi-window predictions, enabling multi-shape kernel interactions to capture cross-modality correspondence; 2) a Fuse-and-Refine Transformer to aggregate the windowed outputs and learn residuals for detail refinement; and 3) a joint adversarial training objective to enhance window-conditioned realism. Extensive experiments demonstrate that our compact WING achieves state-of-the-art performance on the MRI-to-CT and CBCT-to-CT benchmarks, while supporting multi-anatomy synthesis with a single model.
