Decoupling Respiratory and Angular Variation in Rotational X-ray Scans Using a Prior Bilinear Model

Tobias Geimer, Paul Keall, Katharina Breininger, Vincent Caillet, Michelle Dunbar, Christoph Bert, Andreas MaierarXiv 2018 · arXiv

TL;DR

This paper introduces a novel way to separate respiratory motion from gantry rotation in X-ray projections using a bilinear model derived from a patient's prior 4D CT scan. By leveraging the known gantry angle to interpolate rotational weights via B-splines, the authors transform a complex, ill-posed decoupling problem into a simple linear estimation of respiratory state, enabling more accurate motion tracking during radiation therapy.

What Problem Does This Paper Address?

In Image-Guided Radiation Therapy (IGRT), clinicians face the challenge of "hitting a moving target." As a patient breathes, internal organs and tumors shift. To compensate for this, doctors often use rotational X-ray scans (like Cone-Beam CT) during treatment to monitor the patient's state.

However, these scans present a "signal mixing" problem: the appearance of the X-ray projection changes simultaneously because the X-ray source is rotating around the patient (angular variation) and because the patient is breathing (respiratory variation). Standard data-driven methods often struggle to disentangle these two effects, especially when the gantry is moving continuously. Most existing solutions either only extract a simplistic 1D breathing signal or are restricted to static viewing angles, making them ill-suited for modern volumetric arc therapy.

Methods

The authors propose a bilinear model to decouple these variations, grounded in the mathematical linearity of the X-ray transform.

  1. Tensor Construction: They start with a prior 4D CT scan of the patient. From this, they generate Digitally Reconstructed Radiographs (DRRs) at various angles and respiratory phases to create a large data tensor $\mathcal{D}$ (Pixels $\times$ Phases $\times$ Angles).
  2. Dimensionality Reduction: Using Higher-order Singular Value Decomposition (HOSVD), they decompose this tensor into a core "model tensor" $\mathcal{M}$ and two sets of low-dimensional weights: one for respiration ($A$) and one for rotation ($B$).

    Fig. 1. Left: Tensor unfolding concatenates the slices of a tensor along a selected mode. Right: Dimensionality reduction (HOSVD) of the rotational and respiratory domain.
  3. The B-Spline Trick: Estimating both rotation and respiration weights from a single new X-ray image is normally an "ill-posed" problem (too many unknowns). However, the authors realize that in a real clinical setting, the gantry angle is known. They use B-spline interpolation to calculate the rotational weights for any given angle based on the training set.

  4. Linear Estimation: Once the rotational weights are fixed, finding the respiratory weights becomes a straightforward linear problem that can be solved using a pseudo-inverse. These weights can then drive a 3D motion model to predict the tumor's exact position in real-time.

Datasets

The study utilized 4D CT datasets from 5 patients. Each dataset consisted of eight respiratory phases (from full inhale to full exhale).

  • Training/Testing: They generated DRRs (512x512 pixels) at 60 different angles (6° steps).
  • Validation: They used a "leave-one-phase-out" and "leave-some-angles-out" approach to ensure the model could generalize to unseen breathing states and viewing angles.

Key Insights and Ideas

  • Physics-Based Modeling: The core "aha!" moment is that the X-ray transform is linear. This means a bilinear model isn't just a statistical convenience (like it is for separating "pose" and "identity" in face recognition); it is a mathematically sound representation of the underlying physics.
  • Mean-Centering Matters: Unlike standard PCA, the bilinear model here does not operate on mean-subtracted data. Consequently, the first principal component effectively represents the "mean" anatomy, while subsequent components capture the actual motion.
  • Feature Correspondence: The authors found that the $(n+1)$-th bilinear respiratory component in the 2D projection space corresponds directly to the $n$-th linear component of the 3D 4D-CT volume. This confirms that the 2D weights successfully capture 3D volumetric motion.
Fig. 2. Feature weights for the first few principal components for linear PCA on the 4D CT (t. l.), bilinear respiratory (b. l.) and rotational weights (b. r.) as well as the explained variance caused by respiration in the linear and bilinear case (t. r.).

Key Results

  • Image Accuracy: The model achieved a mean estimation error of only 3.01% in gray values for unseen viewing angles.
  • Reconstruction Quality: In a leave-one-phase-out test, the average error was $1.28 \pm 1.27$ compared to a reference mean gray value of ~44.
  • 3D Motion Prediction: When using the extracted weights to drive a 3D motion model, the mean voxel error ranged from 25 to 100 Hounsfield Units (HU). Given that HU ranges from -1000 to 3000, this represents highly accurate 3D volume estimation from 2D projections.

Limitations and Future Work

The authors acknowledge two primary hurdles for clinical translation:

  1. Prior Data Requirement: The method requires a high-quality prior 4D CT. While common in radiotherapy, it might not be available in other X-ray applications.
  2. Static Anatomy Assumption: The current model assumes the patient's baseline anatomy doesn't change between the prior CT and the treatment (e.g., no significant weight loss or tumor shrinkage).
  3. Registration: It assumes perfect alignment between the training CT and the treatment imaging system.

Future work will focus on investigating how anatomical changes interfere with the decomposition and exploring the potential to predict entire dense deformation fields rather than just HU-based models.

Summary updated

Abstract

Data-driven respiratory signal extraction from rotational X-ray scans is a challenge as angular effects overlap with respiration-induced change in the scene. In this paper, we use the linearity of the X-ray transform to propose a bilinear model based on a prior 4D scan to separate angular and respiratory variation. The bilinear estimation process is supported by a B-spline interpolation using prior knowledge about the trajectory angle. Consequently, extraction of respiratory features simplifies to a linear problem. Though the need for a prior 4D CT seems steep, our proposed use-case of driving a respiratory motion model in radiation therapy usually meets this requirement. We evaluate on DRRs of 5 patient 4D CTs in a leave-one-phase-out manner and achieve a mean estimation error of 3.01 % in the gray values for unseen viewing angles. We further demonstrate suitability of the extracted weights to drive a motion model for treatments with a continuously rotating gantry.
BibTeX
@misc{geimer2018decoupling,
  author = {Tobias Geimer and Paul Keall and Katharina Breininger and Vincent Caillet and Michelle Dunbar and Christoph Bert and Andreas Maier},
  title = {{Decoupling Respiratory and Angular Variation in Rotational X-ray Scans Using a Prior Bilinear Model}},
  year = {2018},
  eprint = {1804.11227},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1804.11227},
}

Loading related content...

Figure