[Paper Review] Comparison of reconstruction algorithms for digital breast tomosynthesis

Ingrid Reiser, J Bian+3 more|ArXiv.org|Aug 18, 2009

Digital Radiography and Breast ImagingMedicine11 references21 citations

TL;DR

This study systematically evaluates three digital breast tomosynthesis (DBT) reconstruction algorithms—filtered backprojection (FBP), expectation maximization (EM), and total variation (TV)-minimization—using a clinically relevant digital breast phantom. Results show that iterative methods, especially TV-minimization, reduce artifacts and improve in-depth resolution compared to FBP, which exhibits higher noise and streaking artifacts due to ramp filtering.

ABSTRACT

Digital breast tomosynthesis (DBT) is an emerging modality for breast imaging. A typical tomosynthesis image is reconstructed from projection data acquired at a limited number of views over a limited angular range. In general, the quantitative accuracy of the image can be significantly compromised by severe artifacts and non-isotropic resolution resulting from the incomplete data. Nevertheless, it has been demonstrated that DBT may yield useful information for detection/classification tasks and thus is considered a promising breast imaging modality currently undergoing pre-clinical evaluation trials. The purpose of this work is to conduct a preliminary, but systematic, investigation and evaluation of the properties of reconstruction algorithms that have been proposed for DBT. We use a breast phantom designed for DBT evaluation to generate analytic projection data for a typical DBT configuration, which is currently undergoing pre-clinical evaluation. The reconstruction algorithms under comparison include (i) filtered backprojection (FBP), (ii) expectation maximization (EM), and (iii) TV-minimization algorithms. Results of our studies indicate that FBP reconstructed images are generally noisier and demonstrate lower in-depth resolution than those obtained through iterative reconstruction and that the TV-minimization reconstruction yield images with reduced artifacts as compared to that obtained with other algorithms under study.

Motivation & Objective

• To evaluate and compare the performance of major DBT reconstruction algorithms under realistic imaging conditions.
• To assess image quality metrics such as artifact reduction, resolution, noise characteristics, and uniformity in reconstructed volumes.
• To investigate the impact of data incompleteness on image fidelity using a phantom designed for DBT evaluation.
• To determine whether iterative reconstruction methods like EM and TV-minimization offer advantages over traditional FBP in clinical-relevant tasks.
• To provide a systematic benchmark for algorithm selection in pre-clinical DBT development.

Proposed method

• A digital breast phantom was designed with anatomically relevant structures, including fibroglandular tissue, tumors, microcalcifications, and muscle, to simulate clinical DBT conditions.
• Analytic projection data were generated from the phantom for a typical clinical DBT acquisition geometry with 11–21 views over a 15°–50° angular range.
• Three reconstruction algorithms were applied: filtered backprojection (FBP), expectation maximization (EM), and total variation (TV)-minimization.
• Image quality was evaluated using 2D and 1D profile analysis, visual inspection of reconstructed slices, and region-of-interest (ROI) comparisons in in-focus planes.
• Noiseless and noisy data were used to assess the effects of quantum noise and structure noise on reconstruction fidelity.
• Key metrics included in-plane and in-depth resolution, artifact levels (e.g., streaking), contrast preservation, and uniformity of reconstructed structures.

Experimental results

Research questions

• RQ1 How do FBP, EM, and TV-minimization algorithms compare in terms of artifact suppression in DBT reconstructions?
• RQ2 To what extent do iterative reconstruction methods improve in-depth resolution compared to FBP in limited-angle DBT?
• RQ3 How do noise characteristics (quantum vs. structure noise) differ across reconstruction algorithms in DBT?
• RQ4 Can TV-minimization effectively preserve the shape and contrast of objects with varying in-plane and in-depth profiles?
• RQ5 How well do the algorithms reconstruct microcalcification clusters, a key diagnostic feature in breast cancer?

Key findings

• FBP reconstructions exhibited significantly higher quantum noise compared to iterative methods, with pronounced streaking artifacts due to ramp filtering.
• Iterative methods (EM and TV) demonstrated superior in-depth resolution, reducing ghosting and overlapping artifacts, especially for stacked spheres spaced by 2d.
• TV-minimization produced the smoothest images with reduced artifacts, though it introduced subtle 'speckle noise' or spikes in otherwise uniform regions.
• The EM algorithm showed a noticeable DC shift and increased intensity in uniform regions, particularly in the pectoralis muscle, compared to the true phantom.
• In microcalcification cluster reconstructions, TV-minimization outperformed FBP and EM by minimizing streaking artifacts, especially in noisy conditions.
• Profile analysis confirmed that in-plane resolution was well-preserved across all methods, but in-depth resolution was significantly improved with iterative algorithms.