Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DuoLift-GAN:Reconstructing CT from Single-view and Biplanar X-Rays with Generative Adversarial Networks

Zhaoxi Zhang, Yueliang Ying

TL;DR

DuoLift-GAN tackles the problem of constructing 3D chest CT volumes from limited 2D X-ray data by introducing dual lifting branches that elevate 2D projections and features into 3D representations, merged into a unified 3D feature map and decoded into a full volume. A masked loss focuses learning on intra-lung regions to improve structural detail, while a discriminative adversarial component encourages textural realism. On the LIDC-IDRI dataset, DuoLift-GAN and its CNN variant outperform state-of-the-art single- and double-view methods across PSNR, SSIM, and LPIPS, with ROIs such as vessels being enhanced by the 3D-texture emphasis. The work highlights a trade-off between perceptual detail and traditional pixel- or structure-based metrics and offers a practical approach for intraoperative or resource-limited contexts, with code and weights planned for release to support further research.

Abstract

Computed tomography (CT) provides highly detailed three-dimensional (3D) medical images but is costly, time-consuming, and often inaccessible in intraoperative settings (Organization et al. 2011). Recent advancements have explored reconstructing 3D chest volumes from sparse 2D X-rays, such as single-view or orthogonal double-view images. However, current models tend to process 2D images in a planar manner, prioritizing visual realism over structural accuracy. In this work, we introduce DuoLift Generative Adversarial Networks (DuoLift-GAN), a novel architecture with dual branches that independently elevate 2D images and their features into 3D representations. These 3D outputs are merged into a unified 3D feature map and decoded into a complete 3D chest volume, enabling richer 3D information capture. We also present a masked loss function that directs reconstruction towards critical anatomical regions, improving structural accuracy and visual quality. This paper demonstrates that DuoLift-GAN significantly enhances reconstruction accuracy while achieving superior visual realism compared to existing methods.

DuoLift-GAN:Reconstructing CT from Single-view and Biplanar X-Rays with Generative Adversarial Networks

TL;DR

DuoLift-GAN tackles the problem of constructing 3D chest CT volumes from limited 2D X-ray data by introducing dual lifting branches that elevate 2D projections and features into 3D representations, merged into a unified 3D feature map and decoded into a full volume. A masked loss focuses learning on intra-lung regions to improve structural detail, while a discriminative adversarial component encourages textural realism. On the LIDC-IDRI dataset, DuoLift-GAN and its CNN variant outperform state-of-the-art single- and double-view methods across PSNR, SSIM, and LPIPS, with ROIs such as vessels being enhanced by the 3D-texture emphasis. The work highlights a trade-off between perceptual detail and traditional pixel- or structure-based metrics and offers a practical approach for intraoperative or resource-limited contexts, with code and weights planned for release to support further research.

Abstract

Computed tomography (CT) provides highly detailed three-dimensional (3D) medical images but is costly, time-consuming, and often inaccessible in intraoperative settings (Organization et al. 2011). Recent advancements have explored reconstructing 3D chest volumes from sparse 2D X-rays, such as single-view or orthogonal double-view images. However, current models tend to process 2D images in a planar manner, prioritizing visual realism over structural accuracy. In this work, we introduce DuoLift Generative Adversarial Networks (DuoLift-GAN), a novel architecture with dual branches that independently elevate 2D images and their features into 3D representations. These 3D outputs are merged into a unified 3D feature map and decoded into a complete 3D chest volume, enabling richer 3D information capture. We also present a masked loss function that directs reconstruction towards critical anatomical regions, improving structural accuracy and visual quality. This paper demonstrates that DuoLift-GAN significantly enhances reconstruction accuracy while achieving superior visual realism compared to existing methods.

Paper Structure

This paper contains 27 sections, 8 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. CT reconstruction from single or double view X-rays
  4. Chest CT Reconstruction from X-rays
  5. Reconstruction Quality Evaluation
  6. Method
  7. Generator Architecture
  8. Discriminator Architecture
  9. Train
  10. Loss
  11. Experiments
  12. Dataset Description
  13. LIDC-IDRI
  14. Evaluation Metrics
  15. Structural Similarity Index Measure (SSIM)
  16. ...and 12 more sections

Figures (2)

  • Figure 1: Network architecture of DuoLift-GAN. From left to right, DRRs produced through forward projection on target volume are passed into the Generator to generate the volumetric reconstruction. The generated reconstruction and target volume are then input into the Discriminator.
  • Figure 2: The visualization of reconstructed volumes from the chosen sample in the test set of the LIDC-IDRI dataset, grouped by the corresponding model. The left side shows 3D renderings, while the right side presents the reconstructed volume's coronal, axial, and sagittal slices.