💻 computer science

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

This paper introduces WING, a novel generative network that leverages a window-prior strategy and a Gated Inception architecture to overcome the challenges of high dynamic range in CT synthesis, achieving state-of-the-art performance in cross-modality MRI-to-CT and CBCT-to-CT conversion for adaptive radiotherapy.

Original authors: Siyuan Mei, Yan Xia, Yipeng Sun, Siming Bayer, Zirong Li, Chengze Ye, Daiqi Liu, Fuxin Fan, Yixing Huang, Andreas Maier

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Siyuan Mei, Yan Xia, Yipeng Sun, Siming Bayer, Zirong Li, Chengze Ye, Daiqi Liu, Fuxin Fan, Yixing Huang, Andreas Maier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to recreate a complex, high-definition 3D map of a city (a CT scan) using only a blurry, black-and-white sketch (an MRI or a low-radiation CBCT). This is a common challenge in radiation therapy: doctors need the precise "density map" of a CT scan to calculate radiation doses, but they want to avoid giving patients extra radiation just to get that map.

The problem is that CT scans are like a city with extreme contrasts: some parts are as dark as a deep cave (air in lungs), some are as bright as a neon sign (dense bone), and most are in the middle (soft tissue). If you try to teach a computer to draw the whole city at once, it gets overwhelmed. It tends to focus on the "middle" parts (soft tissue) because there are so many of them, and it accidentally blurs out the important extremes (the caves and the neon signs).

The authors of this paper, WING, decided to change the strategy. Instead of asking the computer to draw the whole city in one go, they asked it to draw the city in three different "zoomed-in" views at the same time, then stitch them back together.

Here is how their solution works, broken down into simple concepts:

1. The "Window" Strategy (The Three Lenses)

Think of a CT scan like a photo that needs different filters to see different things clearly.

The Lung Window: A filter that makes the air and soft tissue visible but ignores the bright bones.
The Soft Tissue Window: A filter that highlights organs and tumors but ignores the air and the super-dense bones.
The Bone Window: A filter that makes the skeleton pop out while ignoring the rest.

The paper argues that these three views are much easier for a computer to learn than the full, chaotic image. It's like asking an artist to draw a portrait by first sketching just the eyes, then just the mouth, and then just the hair, rather than trying to get every detail perfect in one single stroke.

2. The "Gated Inception" Generator (The Specialized Art Team)

To create these three views, the authors built a new type of AI brain called a Gated Inception Generator.

The Analogy: Imagine a team of artists where each artist specializes in a different shape. One is great at drawing long, thin lines (like ribs), another is great at wide, flat shapes (like lungs), and another is good at small details.
How it works: Instead of using one giant brush, this network splits the job into four smaller "branches" (like different artists). They look at the input image through different "lenses" (kernels) to catch details from different angles.
The Gate: A smart "manager" (the Gated part) decides how much to listen to each artist. If the image needs more bone detail, the manager turns up the volume on the "bone artist" and turns down the others. This ensures the AI captures the specific shapes needed for each view.

3. The "Fuse-and-Refine" Transformer (The Master Editor)

Once the AI has drawn the three separate views (Lung, Soft Tissue, Bone), it needs to combine them into one final image.

The Problem: If you just tape three photos together, you might see ugly seams where they overlap.
The Solution: The authors use a Fuse-and-Refine Transformer. Think of this as a master editor who takes the three drafts and blends them together smoothly.
The "Residual" Trick: The editor doesn't just blend them; it looks at the blended result, finds the mistakes (the "residuals"), and adds a final layer of polish to fix the details. This ensures the final image looks sharp and realistic, not blurry at the seams.

4. The "Judge" (Adversarial Training)

Finally, the system uses a "Judge" (a discriminator) that acts like a strict art critic.

The critic looks at the AI's three separate drafts and the final blended image.
It compares them to real, perfect CT scans.
If the AI's work looks fake or blurry, the critic sends it back to the drawing board. This forces the AI to keep improving until the synthetic CT looks indistinguishable from a real one.

The Result

The paper claims that by using this "Window" approach, their model (WING) is better at recreating the tricky parts of the body—like the lungs and bones—than previous methods. It works well whether the starting image is an MRI or a low-radiation CBCT.

In short, instead of trying to solve a giant, messy puzzle all at once, WING breaks the puzzle into three manageable sections, solves each one with a specialized tool, and then expertly glues them back together to create a perfect picture.

Technical Summary: WING – A Window-Prior-Based Generative Network for Cross-Modality CT Synthesis

1. Problem Statement

In adaptive radiotherapy (ART), accurate treatment planning requires CT-calibrated X-ray attenuation maps (Hounsfield Units, HU) for dose calculation. However, workflows relying solely on MRI or Cone-Beam CT (CBCT) lack these specific density values, necessitating the generation of synthetic CT (sCT).

Current deep learning approaches for sCT generation typically treat the task as a direct 3D image-to-image regression problem. The authors identify a fundamental limitation in this approach: CT intensities possess an absolute, absolute dynamic range (e.g., -1000 to 3000 HU) with highly imbalanced, long-tailed distributions. Direct regression is dominated by abundant soft-tissue voxels (typically [-200, 200] HU), causing the model to "average out" sparse but clinically critical structures such as lung parenchyma (low density) and bone (high density). While some methods attempt to mitigate this with geometric guidance, they often require impractical external segmentation labels.

2. Methodology

The paper proposes WING (a WINdow-prior-based Generative Network), which reformulates the regression target from full-range CT to multiple windowed representations. This leverages the inductive prior that CT intensities are structure-deterministic and window-separable.

Core Components:

CT Windowing and Soft Fusion:
- Windowing: Instead of predicting raw HU values, the network predicts three non-overlapping windowed views corresponding to clinical standards: Lung (WL: -600, WW: 800), Soft Tissue (WL: 0, WW: 400), and Bone (WL: 1100, WW: 1800). These views exhibit smoother distributions and clearer structural contrasts.
- Soft Fusion: To reconstruct the full-range CT from the windowed predictions, the authors introduce a differentiable soft fusion function. Rather than hard boundaries, weighted masks are derived from the soft-tissue window to blend the denormalized lung, soft tissue, and bone predictions smoothly. A parameter $\tau$ controls the transition sharpness.
Gated Inception Generator (GIG):
- Built upon the compact MedNeXt backbone, the generator replaces standard large-kernel 3D depthwise convolutions (which are computationally prohibitive for volumetric data) with a Gated Inception Block (GIB).
- The GIB splits input channels into four branches: one identity branch and three branches using anisotropic kernels ( $3\times3\times3$ , $3\times9\times1$ , $3\times1\times9$ ).
- A gated fusion mechanism (using a lightweight point-wise convolution and Sigmoid) adaptively reweights these branches, allowing the network to capture multi-directional spatial information and preserve thin, high-contrast structures without inter-slice blurring.
Fuse-and-Refine Transformer (FRT):
- Recognizing that direct fusion of windowed outputs may lack fine detail, the FRT aggregates the synthesized windowed views and learns residuals to refine the structure.
- The windowed predictions are projected into patch embeddings and processed by Transformer blocks to capture global dependencies across the different window representations.
- The output residual is added to the coarse fused CT to produce the final refined sCT.
Joint Adversarial Objective:
- The training employs a Joint Discriminator (JDisc) conditioned on both the windowed representations and the full-range CT.
- The adversarial loss is defined to simultaneously optimize the realism of the windowed predictions and the fused full-range CT, ensuring dual-level perceptual enhancement.
- The total loss combines $\ell_1$ reconstruction losses (for both windowed and fused outputs) and an LPIPS loss for perceptual quality.

3. Key Contributions

Reformulation of Regression Target: The paper shifts the sCT generation paradigm from direct full-range regression to a multi-windowed approach, explicitly utilizing the structure determinism and window separability of CT as inductive priors to address long-tail distribution issues.
Novel Architecture (WING):
- Introduction of the Gated Inception Generator (GIG) for efficient, multi-shape kernel interactions.
- Development of the Fuse-and-Refine Transformer (FRT) to aggregate windowed outputs and refine structural details via residual learning.
Joint Training Strategy: A joint adversarial objective that synchronizes the optimization of window-conditioned and full-range representations.
Single-Model Multi-Anatomy Synthesis: The framework supports synthesis across head-and-neck, thoracic, and abdominal regions using a single compact model.

4. Experimental Results

The method was evaluated on the SynthRAD2025 dataset, containing paired MRI-to-CT and CBCT-to-CT volumes from head-and-neck, thoracic, and abdominal cancer patients.

Performance: WING achieved state-of-the-art (SoTA) performance on both tasks, outperforming previous benchmarks including Pix2Pix3D, Ea-GAN, ResUNet, SwinUNETR, and the MedNeXt baseline.
- MRI-to-CT: Improved MAE by 1.58 HU, MS-SSIM by 0.51, PSNR by 0.40, and DICE by 0.31 over the MedNeXt baseline.
- CBCT-to-CT: Improved MAE by 1.99 HU, MS-SSIM by 0.62, PSNR by 0.26, and DICE by 0.49 over the MedNeXt baseline.
Per-Window Analysis: The model demonstrated significant gains in per-window Mean Absolute Error (MAE), particularly for lung and bone structures (the long-tail extremes), confirming that the windowed approach improves fidelity across heterogeneous attenuation regimes.
Efficiency: Despite the added complexity of the Transformer and Gated Inception blocks, the model remains parameter-efficient (19.29M parameters), comparable to the MedNeXt baseline (17.47M).

5. Significance and Claims

The paper claims that WING is the first window-prior-based generative framework for cross-modality CT synthesis. Its primary significance lies in alleviating the challenge of regressing long-tailed CT intensity distributions without relying on external geometric guidance or segmentation labels.

The authors assert that by explicitly incorporating imaging-driven window priors, WING generates synthetic CT images with high fidelity and stable intensity characteristics, making it suitable for reliable use in radiotherapy workflows. The work demonstrates that decomposing the regression task into windowed views and reassembling them via a learned fusion mechanism is a more effective strategy for preserving clinically important, sparse structures than direct global regression.

Limitations Acknowledged:
The authors note that the predefined window settings follow standard radiological practice, which may limit flexibility under uncommon acquisition protocols. Additionally, the separation of generation and fusion stages introduces design complexity. Future work is suggested to focus on more unified and adaptive window modeling strategies.