WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis

Jul 16, 2026

Upload by Riley

Generating CT volumes from MRI and CBCT enables improved treatment planning in adaptive radiotherapy while reducing additional radiation exposure. Direct CT intensity regression is difficult due to CT’s high dynamic range and long-tailed distributions, which can average out sparse but clinically important anatomical structures. WING reformulates the regression target into multiple windowed representations using the inductive prior that CT intensities are structure-deterministic and window-separable. A Gated Inception Generator produces multi-window outputs, a Fuse-and-Refine Transformer aggregates them for residual-based refinement, and joint adversarial training improves realism. Experiments show strong MRI-to-CT and CBCT-to-CT performance and multi-anatomy synthesis with a single model.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages11

Reading Time28 min

Introduction

Motivation: limitations of direct intensity regression

Synthetic CT in adaptive radiotherapy

Proposed Method

Window-prior reformulation of the regression target

Gated Inception Generator (multi-window predictions)

Fuse-and-Refine Transformer (aggregation and refinement)

Joint adversarial training objective

Experiments (overview)

MRI-to-CT and CBCT-to-CT benchmarks

arXiv :2607 .06234v 1 [ cs .CV] 7 Jul 2026 WING: A Window-Prior-Based Generative Network with Gated Inception for Cross-Modality CT Synthesis Siyuan Mei 1 , Yan Xia2 , Yipeng Sun 1 , Siming Bayer 1 , Zirong Li3 , Chengze Ye 1 , Daiqi Liu 1 , Fuxin Fan3 , Yixing Huang4 , and Andreas Maier 1 1 Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen 91058, Germany [siyuan.mei@fau.de](siyuan.mei@fau.de) 2 Department of Orthodontics and Orofacial Orthopaedics, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen 91054, Germany 3 Digital Technology and Innovation, Siemens Healthineers, Shanghai 201318, China 4 Institute of Medical Technology, Peking University, Beijing 100191, China Abstract. 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-theart performance on the MRI-to-CT and CBCT-to-CT benchmarks, while supporting multi-anatomy synthesis with a single model. Keywords: CT synthesis · CT windows · Gated inception · GAN. 1 Introduction Adaptive radiotherapy (ART) leverages repeated on-treatment imaging, most commonly cone-beam computed tomography (CBCT) on conventional linacsand magnetic resonance imaging (MRI) on MR-guided systems, to capture interfraction anatomical variability and update treatment plans [2,23,21] . While MRI provides excellent soft-tissue contrast and CBCT involves lower radiation exposure, neither modality directly yields CT-calibrated X-ray attenuation required 2 S. Mei et al. Fig. 1. Illustration of raw CT images and three non-overlapping windows of lung, soft tissue, and bone. The windowed views are obtained using the fixed window level (WL) and window width (WW), featuring more learnable distributions. for accurate proton and photon dose calculation [4,24] . To optimize treatment planning in MRI- or CBCT-only ART workflows, synthetic CT (sCT) has recently been introduced to provide CT-equivalent electron density information, thereby avoiding additional diagnostic CT acquisitions and associated patient radiation burden [22,23,8] . A wide range of deep learning models have been applied to sCT generation by casting it as a 3D image-to-image translation problem [22,5,8] . Early approaches mainly rely on convolutional neural networks (CNNs), where U-Net [18] and its variants [9,26] are widely adopted due to their efficient encoder-decoder structures. To further improve image fidelity, subsequent studies introduce advanced generative paradigms, including adversarial generative models (GAN) with discriminators [10,27,30] and diffusion models with iterative denoising processes [15,3] . More recently, higher-capacity modern architectures such as Transformers [1] and ConvNeXts [19,12] have also been explored to capture richer spatial representations with long-range dependencies. Despite their architectural diversity, these methods overlook CT-s

FAQs

Why is direct full-range regression of CT intensities challenging for MRI-to-CT or CBCT-to-CT synthesis?

CT intensities have a much wider dynamic range and highly imbalanced long-tailed distributions, so learning can be dominated by abundant soft-tissue voxels and suppress long-tail structures that are clinically important.

How does WING reformulate the CT synthesis learning target?

WING shifts the regression target from the full-range CT to multiple windowed representations, leveraging that CT intensities are structure-deterministic and window-separable, which yields smoother distributions and structured fusion back to full-range CT.

What components does WING use to generate and refine CT outputs from multiple windows?

WING introduces a Gated Inception Generator to produce multi-window predictions, a Fuse-and-Refine Transformer to aggregate windowed outputs and learn residual refinements, and a joint adversarial training objective to enhance window-conditioned realism.

Jul 16, 2026

Upload by Riley