Numerical Methods for Coupled Reconstruction and Registration in Digital Breast Tomosynthesis

Guang Yang, John H. Hipwell, David J. Hawkes, Simon R. Arridge

Abstract

In medical imaging, Digital Breast Tomosynthesis (DBT) often faces challenges due to its limited-angle acquisition, which results in ill-posed reconstruction problems and significant out-of-plane artifacts. Traditional workflows typically treat 3D reconstruction and temporal registration as sequential, independent tasks. This paper introduces a novel mathematical framework that performs these tasks simultaneously. By coupling reconstruction and registration into a single objective function, the proposed method significantly reduces artifacts and improves spatial alignment accuracy compared to traditional sequential approaches.

Introduction

Digital Breast Tomosynthesis (DBT) has become a vital tool in clinical radiology, offering a 3D perspective that helps identify lesions obscured by overlapping tissue in standard 2D mammography. However, unlike Computed Tomography (CT), DBT acquisitions are restricted to a narrow angular range (e.g., ±25°). This limitation makes the reconstruction process inherently "ill-posed," leading to blurring and artifacts that can hinder diagnostic accuracy.

A common clinical requirement is temporal registration—aligning a current scan with a previous one to monitor changes such as tumor growth. The standard "reconstruct then register" workflow is often suboptimal because reconstruction artifacts can mislead registration algorithms, while poor registration prevents the effective use of prior data to refine the reconstruction. This paper explores a joint framework to bridge this gap.

The Challenge of DBT Reconstruction

The primary difficulty in DBT lies in its "null space"—the missing data resulting from the limited projection angles. When radiologists attempt to compare volumes over time, the artifacts generated during the initial reconstruction phase often propagate into the registration phase. Conversely, if the registration is inaccurate, it cannot provide the necessary regularization to fill the gaps in the reconstruction's missing data.

Methodology

The researchers propose moving beyond sequential methods toward a fully coupled approach. They define a single objective function that incorporates both image intensities (reconstruction) and transformation parameters (registration).

Optimization Strategies

Two primary approaches were evaluated:

  1. Iterative Method (Partially Coupled): This method alternates between incomplete reconstruction steps and registration steps, using the current registration estimate to "correct" the reconstruction in a loop.
  2. Simultaneous Method (Fully Coupled): This core innovation minimizes a joint objective function using L-BFGS (a quasi-Newton optimization algorithm) to solve for image intensities and transformation parameters (including rotation, scaling, and non-rigid warping) simultaneously.

The framework supports both Affine transformations for global alignment and non-rigid B-spline models for local deformations, allowing it to handle the complex physical compressions seen in clinical breast imaging.

Comparison of the proposed DBT model with existing coupled reconstruction and registration methods across various imaging modalities.

As shown in the table above, the proposed model for DBT utilizes 3D Affine and B-spline dimensions with L-BFGS or Conjugate Gradient optimizers, placing it within a broader context of coupled methods used in fields like PET, MRI, and Cryo-EM.

Experimental Validation

The methods were validated using three distinct datasets to ensure robustness:

  • Software Synthetic Toroidal Phantom: A controlled geometric environment used to test 20 random affine transformations.
  • Uncompressed Breast MR Images: High-resolution 3D MRI data used to simulate 2D DBT projections, providing a "ground truth" for fidelity assessment.
  • In-vivo DBT Simulations: Real-world MRI acquisitions of breasts under different physical plate compressions, mimicking clinical non-rigid deformations.

Key Results and Insights

The quantitative improvements demonstrated by the simultaneous method were substantial:

  • Reconstruction Fidelity: In the Toroid phantom tests, the Simultaneous Method achieved a Mean Squared Error (MSE) of $0.24 \times 10^3$, a nearly 30-fold improvement over the Sequential Method's $7.46 \times 10^3$.
  • Registration Accuracy: In Breast MRI simulations, the relative error for the Simultaneous Method (0.0026) was approximately half that of the Sequential Method (0.0051).
  • Artifact Suppression: Qualitatively, the joint framework effectively distinguished between actual tissue structures and reconstruction artifacts. By using registration as a form of regularization, the algorithm "filled in" the null space, resulting in sharper images with significantly reduced out-of-plane blur.

Conclusion and Future Directions

By treating reconstruction and registration as reciprocal tasks, this framework demonstrates that better registration provides more information for reconstruction, while cleaner reconstructions allow for more precise registration.

Despite these gains, challenges remain regarding computational cost. The authors suggest that GPU acceleration for forward and backward projectors is essential for clinical viability. Future work will also focus on multi-scale optimization to avoid local minima and extending the framework to integrate multiple views (e.g., CC and MLO) into a single high-resolution 3D volume. This research paves the way for more accurate temporal monitoring and improved diagnostic confidence in breast cancer screening.

Summary updated

BibTeX
@inproceedings{yang2013numerical,
  author = {Guang Yang and John H. Hipwell and David J. Hawkes and Simon R. Arridge},
  title = {{Numerical Methods for Coupled Reconstruction and Registration in Digital Breast Tomosynthesis}},
  booktitle = {Spiral},
  year = {2013},
}
Figure