PaperlayerView exact revision on arXiv

Source-linked AI summary

Enhanced imaging of microcalcifications in digital breast tomosynthesis through improved image-reconstruction algorithms

Emil Y. Sidky, Xiaochuan Pan, Ingrid S. Reiser, Robert M. Nishikawa, Richard H. Moore, Daniel B. Kopans

arXiv:0904.1016v1physics.med-ph

TL;DR

Under-sampled DBT reconstruction is nonunique because many varied volumes can satisfy a given data tolerance. The paper develops a practical iterative ASD-POCS-based algorithm using total p-variation and convex constraints, and applies it to three clinical DBT datasets. Results indicate a substantial potential advantage for microcalcification imaging, especially at low p values.

  • Problem

    Under-sampled DBT data can correspond to many widely varied reconstructed volumes, creating a need for reconstruction that handles incomplete projections.

  • Method

    The paper develops a practical iterative ASD-POCS reconstruction algorithm that minimizes total p-variation while enforcing image and projection-data constraints.

  • Results

    Images reconstructed with low p values show markedly greater microcalcification contrast than images from existing algorithms.

  • Takeaways & Limitations

    The results indicate a potential substantial advantage of the proposed reconstruction algorithm for microcalcification imaging in DBT.

Abstract

from arXiv · show

PURPOSE: We develop a practical, iterative algorithm for image-reconstruction in under-sampled tomographic systems, such as digital breast tomosynthesis (DBT). METHOD: The algorithm controls image regularity by minimizing the image total $p$-variation (TpV), a function that reduces to the total variation when $p=1.0$ or the image roughness when $p=2.0$. Constraints on the image, such as image positivity and estimated projection-data tolerance, are enforced by projection onto convex sets (POCS). The fact that the tomographic system is under-sampled translates to the mathematical property that many widely varied resultant volumes may correspond to a given data tolerance. Thus the application of image regularity serves two purposes: (1) reduction of the number of resultant volumes out of those allowed by fixing the data tolerance, finding the minimum image TpV for fixed data tolerance, and (2) traditional regularization, sacrificing data fidelity for higher image regularity. The present algorithm allows for this dual role of image regularity in under-sampled tomography. RESULTS: The proposed image-reconstruction algorithm is applied to three clinical DBT data sets. The DBT cases include one with microcalcifications and two with masses. CONCLUSION: Results indicate that there may be a substantial advantage in using the present image-reconstruction algorithm for microcalcification imaging.

I. INTRODUCTION

DBT uses incomplete projection data, making reconstruction nonunique and challenging despite relatively high-quality measurements. The paper adapts and simplifies the ASD-POCS framework to provide practical iterative reconstruction with image regularity and physical constraints.

  • Motivation: DBT projection data are high quality but radically incomplete, so many different attenuation distributions can agree with the available measurements.This differs from systems designed for complete or nearly complete data that are mainly degraded by noise.
  • Contribution: The proposed approach adapts a compressive-sensing reconstruction algorithm to DBT within the ASD-POCS framework.The framework combines projection-based data and physical constraints with regularization minimization.
  • Motivation: Incomplete projections give regularization two roles: selecting an image among data-consistent candidates and regularizing while relaxing data consistency.The first role minimizes image regularity at fixed data tolerance; the second permits further regularization by relaxing the data constraint.
  • Contribution: The paper reassembles ASD-POCS components into a simplified practical algorithm intended to produce useful images within 10-20 iterations.Simplification reduces algorithm parameters to those with significant effects during the first few iterations.
  • Evaluation and scope: The method is demonstrated on clinical DBT cases containing microcalcifications and masses, while the paper makes few quantitative algorithm comparisons.The authors present ASD-POCS as a framework rather than claiming a single optimal algorithm.

II. SYSTEM MODEL AND IMAGE-RECONSTRUCTION

The paper models DBT as an under-sampled, finite linear tomography system in which voxelized attenuation volumes are related to measured projection data. Its geometry and ray-driven system matrix specify the reconstruction problem, while data incompleteness permits multiple image volumes consistent with the measurements.

  • DBT acquires projections over a limited angular range rather than a full circular CT scan.
  • The prototype uses 11 approximately equally spaced projections across a 50° arc and a 1800x2304 detector with 100-micron bins.
  • The discrete model M⃗f = ˜g approximates continuous X-ray line integrals using a finite voxel coefficient vector and projection data.
  • The imaging volume is represented by voxels measuring 0.1x0.1x1.0 mm^3, with 60 slices used for the presented reconstructions.
  • DBT data incompleteness arises because measurements can be fewer than unknowns and the system matrix can be ill-conditioned.
  • The present setup is under-sampled by a factor of 5 based on 110,880,000 voxel unknowns and 22,351,560 measured rays.

III. ITERATIVE ALGORITHMS AND DBT IMAGE-RECONSTRUCTION

Under-sampled DBT admits many image estimates with different data errors and regularity values, so reconstruction should independently control both quantities. Existing iterative approaches motivate regularity-based constrained optimization.

  • For under-sampled systems, multiple candidate volumes may share the minimum data error while differing in image regularity.
  • Effective reconstruction therefore requires independent control of data error and image regularity.
  • Standard fixed-strength regularization can reduce regularity while slowing data-error reduction, making it potentially inefficient for under-sampled tomography.
  • ASD-POCS combines projection-based constraints with a regularity measure, using total variation to seek minimum-TV images for a fixed data-error tolerance.

IV. A PRACTICAL IMAGE-RECONSTRUCTION ALGORITHM USING THE ASD-POCS FRAMEWORK

The paper presents a practical ASD-POCS reconstruction framework that combines convex-constraint projections with steepest descent and adaptive step control. It reduces parameter burden and uses iteration number and relaxation to navigate the data-error–regularity trade-off.

  • POCS enforces convex image constraints, while steepest descent reduces the image regularity measure.
  • The line search maximizes the steepest-descent step while preventing increases in the regularity objective.
  • The revised ASD-POCS algorithm is designed for practicality, with fewer control parameters and a pseudo-code implementation.
  • The adaptive rule permits larger regularity-reduction steps early and constrains them later so data error does not increase.
  • Iteration number controls data-error reduction, whereas lower β reduces the image regularity measure.
  • The authors do not claim ASD-POCS is optimal and describe it as a framework for generating specific reconstruction algorithms.

V. APPLICATION TO DBT PROJECTION DATA

The practical ASD-POCS algorithm is applied to clinical DBT projection data from the GE-MGH instrument. The evaluated cases include one with microcalcifications and two with masses, with a reported significant impact on microcalcification imaging.

  • The study reconstructs clinical DBT projection data acquired on the GE-MGH instrument.
  • Three clinical cases include one case containing microcalcifications and two cases containing masses.
  • The authors report that ASD-POCS can have a significant impact on microcalcification imaging.

A. DBT projection data

The DBT scan uses 11 projection views over a 50° arc. Projections affected by the compression-paddle fin are cropped to reduce edge artifacts and support convergence evaluation.

  • The scan consists of 11 projection views acquired over a 50° arc.
  • A projection at a 25° offset can contain the compression-paddle fin.
  • Projections are truncated to remove rays passing through the fin, which is outside the reconstruction volume.
  • This cropping reduces artifacts at the reconstruction-volume edge and enables demonstration of ASD-POCS convergence properties.

B. Form of the ASD-POCS objective function and algorithm parameters

ASD-POCS uses the total p-variation of the image as a generic objective, with p controlling the regularity measure and image characteristics. The reconstruction uses thick slices because limited angular coverage provides little depth information, while several parameterization choices remain unexplored.

  • For DBT reconstruction, ASD-POCS employs a total p-variation norm as its image objective.
  • When p = 1.0, TpV becomes the convex standard TV norm; when p = 2.0, it becomes a quadratic roughness measure.
  • The value of p significantly affects image quality, and the study reconstructs images with p = 0.8, 1.0, and 2.0.
  • The reconstruction volume contains 60 slices that are 1 mm thick, with 0.1 mm in-plane voxels and voxels ten times longer in depth.
  • Limited DBT angular coverage provides little depth information, motivating the use of thick slices.
  • Thinner slices and spatial differencing may improve depth resolution, but these factors make little difference for ASD-POCS in the 10–20 iteration range.

C. Reconstructed images

The study applies ASD-POCS and a basic EM implementation to three clinical DBT cases, using regularization controls to examine microcalcification and mass reconstructions. ASD-POCS images change little from 5 to 20 iterations, while its parameters are intended for task-specific optimization.

  • Three clinical DBT data sets are reconstructed: one containing microcalcifications and two containing masses.
  • The study primarily demonstrates ASD-POCS image-regularization controls, which can be optimized for different tasks in future work.
  • The cases use the same algorithm-parameter sets, with differences limited mainly to volume dimensions and projection-data cropping.
  • EM and ASD-POCS images are displayed at 5, 10, and 20 iterations, with ROIs showing either microcalcifications or masses.
  • ASD-POCS varies p across 0.8, 1.0, and 2.0, while β takes values 1.0, 0.5, and 0.1.
  • Lower p values tend to sharpen edges, and smaller β generally allows ASD-POCS to reach lower TpV objective values.
  • ASD-POCS reconstructions show surprisingly little change across 5–20 iterations, with practical implications for reconstruction use.

D. Case 1: microcalcifications

In the microcalcification case, ASD-POCS makes these small features more prominent, especially at lower p values, while iteration changes have little visible effect. The authors frame the advantage as potential and task-dependent, with parameter choices requiring further study.

  • ASD-POCS reconstructions prominently display microcalcifications, with lower p values accentuating them more than larger p values.
  • Even at p = 2.0, microcalcification visibility is comparable to the EM results.
  • Microcalcification contrast differences are quantified using profiles along depth and transverse lines intersecting a single microcalcification.
  • ASD-POCS images show little change between 5 and 20 iterations, whereas lower β increases regularization strength.
  • Optimal p and β values for tasks such as human-observer microcalcification detection require separate studies.
  • Data quality also affects parameter selection: lower p may be robust to detector noise but more sensitive to inconsistency.
  • If low-p reconstruction consistently improves microcalcification contrast, it could enable lower probing-beam intensity and reduced DBT radiation dose.

E. Case 2: uniform mass

For the uniform-mass case, ASD-POCS showed weak iteration dependence and parameter-dependent image differences, while mass conspicuity varied less than microcalcification conspicuity. Low p may sharpen edges, but mass-imaging benefits were subtle and require task-based evaluation.

  • ASD-POCS iteration-number dependence appears weak, and mass conspicuity varies less with parameters than microcalcification conspicuity.
  • Lower p sharpens background-feature edges as well as mass edges, so mass conspicuity may not improve dramatically.
  • Similar β-values can produce different apparent image quality across cases; β = 1.0 appears quite noisy here relative to the previous case.
  • β = 0.1 appears visually best for this mass case, whereas β = 0.5 seems best for the previous microcalcification case.
  • For the spiculated mass in dense breast tissue, low-p reconstruction may enhance edges, but any imaging advantage is less clear than for microcalcifications.
  • Mass-imaging advantages need demonstration with task-based image evaluation.
  • Under-regularization at large β tends to yield linear artifacts, while smaller β can wash them out without severely blurring underlying features.

G. Evolution of algorithm metrics

Algorithm trajectories in the data-error–image-regularity plane illustrate how ASD-POCS navigates undersampled DBT reconstruction differently from EM. Lower β can substantially reduce image TV without proportionally increasing data error, although behavior varies across cases and data quality.

  • ASD-POCS aims to reach images across much of the allowed data-error–TV region within about 10 iterations.
  • For the microcalcification case, lowering β directly reduces image TV while adaptive control reduces data error with little change in TV.
  • The minimum data-error image in each sequence occurs at iteration 20, with little dependence on β despite dramatically reduced image TV at lower β.
  • Undersampling permits substantially different-TV images to correspond to similar data error, explaining why regularity can vary without equivalent data-error changes.
  • EM follows the traditional trade-off in which increasing iterations reduces data error at the expense of image regularity.
  • The uniform-mass case shows a significant data-error drop when β decreases from 1.0 to 0.1, because greater regularity can enable further data-error reduction.
  • The spiculated-mass case reaches a minimum data error roughly twice as high as the previous cases, with dense-breast noise and possible geometry errors contributing.
  • Algorithm trajectories may help assess data quality and identify possible patient motion, which is particularly relevant to microcalcification imaging.

VI. DISCUSSION

The paper introduces a practical ASD-POCS reconstruction framework with fine control over image regularity for underdetermined DBT. Across three clinical cases, the clearest benefit is greater microcalcification contrast, while mass-imaging effects are subtler.

  • The paper introduces a practical iterative ASD-POCS image-reconstruction algorithm that produces useful images within a few iterations.
  • The framework provides fine control over reconstructed-image regularity, which is important for underdetermined imaging problems such as DBT.
  • Total p-variation serves as the regularity metric, reducing to total variation at p = 1.0 and image roughness at p = 2.0.
  • The parameters p, β, and iteration number affect reconstructed images, with β controlling the regularity objective-function level.
  • Microcalcification imaging is most strongly impacted: low-p images show markedly greater contrast than images from existing algorithms.
  • Increased microcalcification contrast may permit lower X-ray intensity and therefore lower patient dose for DBT scans.
  • Mass-imaging effects are more subtle, but finer controls may support better optimization for human or computer observers.
  • The framework can be adapted to other X-ray tomographic systems and potentially to linear-data-model imaging modalities.
Loading 0904.1016v1…