Total variation superiorization schemes in proton computed tomography image reconstruction

I. INTRODUCTION

• In iterative pCT reconstruction, image noise increases with iteration number and limits the monotonic improvement of image quality.The introduction also notes that spatial resolution and mean reconstructed RSP improve as iteration number increases.Source: §I · p. 3
• TVS adds perturbations to a feasibility-seeking projection algorithm to steer reconstruction toward feasible images with reduced total variation.The paper distinguishes this goal from constrained minimization: the sought solution need not minimize the merit function.Source: §I · p. 4
• The two investigated schemes apply superiorization steps at different frequencies: once per cycle or once per block.The paper also investigates simplified versions that eliminate the computationally expensive feasibility proximity check.Source: §I · p. 5
• Both TVS-DROP schemes improved image quality relative to standard DROP, with TVS2-DROP yielding the greatest noise reduction and superior density resolution.The introduction also identifies improved detection of small contrast differences as an advantage of TVS.Source: §I · p. 5

A. The superiorization methodology

• The methodology is described as a heuristic, despite relying on the notion of bounded perturbation resilience.The paper presents practical demonstrations of usefulness alongside this theoretical qualification.Source: A. The superiorization methodology · p. 6
• Superiorization perturbs a feasibility-seeking algorithm so it can reduce objective-function values without losing convergence toward feasibility.The method requires a perturbation-resilient algorithm and admissible perturbations that help reduce the objective.Source: A. The superiorization methodology · p. 6
• The schemes differ in perturbation frequency and in when they check the merit function and feasibility proximity.The paper also tests eliminating the computationally expensive feasibility proximity check.Source: A. The superiorization methodology · p. 7

B. Total variation superiorization applied to proton CT

• The pCT system models each proton history with a row of the linear system, where path–voxel intersection lengths weight the unknown image of relative stopping power.The measured-energy-derived integral relative stopping power supplies the corresponding data value.Source: B. Total variation superiorization… · pp. 7–8
• The superiorization paradigm had not been investigated for an empty feasible set, although reducing proximity can still be useful when zero proximity is unattainable.In that case, the proximity function indicates a point that violates the constraints less.Source: B. Total variation superiorization… · pp. 9–10
• The merit function is the reconstructed image's total variation, and perturbations use the negative normalized subgradient to steer estimates toward lower values.The image vector is represented as a two-dimensional image for calculating total variation.Source: B. Total variation superiorization… · p. 10
• TVS1-DROP applies projections across all data blocks before checking proximity using all histories; TVS2-DROP projects and checks using one block at a time.The per-block check is justified by blocks containing equal numbers of histories from each projection angle.Source: B. Total variation superiorization… · pp. 10–12
• The simplified TVS1-DROP* and TVS2-DROP* schemes eliminate the computationally expensive feasibility proximity check.The paper presents this modification as an effort to reduce reconstruction time.Source: B. Total variation superiorization… · p. 13

C. Proton CT Monte Carlo simulations

• Geant4 Monte Carlo simulated the pCT data used to reconstruct two-dimensional images of the phantoms.The phantom set included spatial- and density-resolution phantoms, each described as 16 cm in diameter and containing brain- and cranial-bone-equivalent materials.Source: C. Proton CT Monte Carlo simulations · pp. 13–14
• The calorimeter was modeled with perfect energy resolution, so detector-noise sources were neglected.Source: C. Proton CT Monte Carlo simulations · p. 13
• The simulations used 200 MeV protons, 180 parallel-beam projections at 2-degree intervals, and 20,000 protons per projection angle.The incident protons were monoenergetic.Source: C. Proton CT Monte Carlo simulations · pp. 13–14

• The algorithm variants were compared using spatial resolution, density resolution, and relative RSP accuracy.Source: D. Performance analysis · p. 15
• Spatial resolution was quantified with the 2D modulation transfer function, derived from the point spread function of a dense phantom feature.A 2D FFT of a 16 × 16 pixel region was used, with axial components averaged using phantom symmetry.Source: D. Performance analysis · p. 15
• Low-contrast density resolution was assessed using the contrast discrimination function (CDF), an objective statistical method for determining the minimum contrast required to discriminate an object of a given size from surrounding tissue23.The reconstructed uniform-phantom image was divided into a grid of objects ranging from 1 × 1 to 10 × 10 pixels in size.Source: D. Performance analysis · p. 15
• Relative RSP error was defined to assess reconstructed RSP accuracy, while histogram analysis quantified reconstructed RSP values.Source: D. Performance analysis · pp. 15–16

• Without feasibility proximity checking, the TVS variants had smaller or equivalent minimum relative error compared with checked reconstructions.Source: A. Qualitative comparison · p. 16
• The cycle of minimum relative error was cycle 3 for standard DROP and cycle 10 for TVS1-DROP.Source: A. Qualitative comparison · p. 16
• TVS2-DROP* had the lowest noise level, which the authors suggest may reflect its extra perturbation steps.Source: A. Qualitative comparison · p. 16

• Removing the feasibility proximity check made no difference for TVS1, whose check condition was never violated, but did affect TVS2.TVS1-DROP and TVS1-DROP* were equivalent in quantitative RSP accuracy and had monotonically decreasing relative error.Source: B. Quantitative accuracy · p. 16
• With the feasibility check, TVS2 relative error progressively increased after reaching a minimum similar to standard DROP.The authors attribute this to reduced β dampening the perturbation step’s noise-reducing effect.Source: B. Quantitative accuracy · p. 16
• Minimum relative error was 2.64% for standard DROP, 1.96% for TVS1-DROP and TVS1-DROP*, 1.64% for TVS2-DROP, and 1.55% for TVS2-DROP*.These are minima within the first 10 cycles.Source: B. Quantitative accuracy · p. 17
• All schemes reconstructed the same mean RSP for brain- and bone-equivalent regions within peak-fitting uncertainty, so TVS did not adversely affect those values.Source: B. Quantitative accuracy · p. 17

C. Spatial resolution

• At any spatial frequency, TVS1-DROP* had larger MTF values and better spatial resolution than standard DROP; TVS1-DROP* and TVS2-DROP* performed similarly, with TVS1 marginally better.The authors attribute improved spatial resolution to performing more cycles before reaching the lowest relative error.Source: C. Spatial resolution · p. 17

D. Density resolution

• TVS1-DROP* performed only slightly better than standard DROP on the contrast discrimination function, while TVS2-DROP* performed much better than both.TVS2-DROP* discriminated 1%–1.5% contrast for objects as small as 1 mm2; the authors attribute this to reduced noise and improved spatial resolution.Source: D. Density resolution · p. 18

IV. DISCUSSION

• TVS2-DROP used one perturbation per block iteration, or 12 per cycle here, whereas TVS1-DROP used one perturbation per cycle.The additional perturbation steps in TVS2-DROP were associated with the greatest noise reduction and superior density resolution.Source: §IV · p. 19
• Removing the feasibility proximity check halved reconstruction time and further reduced TVS2-DROP noise without harming other TVS1-DROP performance parameters.The check is intended to prevent TV superiorization from moving the image away from measured data.Source: §IV · p. 19
• Inaccurate forward and backprojection operators make the residual an unreliable guide to image quality amid scattering and energy-straggling data inconsistencies.The authors suggest TVS2-DROP suppresses these inconsistencies.Source: §IV · p. 19
• Both TVS-DROP schemes improved spatial resolution over standard DROP, and TVS2-DROP* retained that advantage despite better noise reduction.The discussion also reports superior density resolution and the greatest noise reduction for TVS2-DROP.Source: §IV · pp. 19–20
• This first quantitative investigation used pCT data from largely uniform virtual phantoms, and realistic anthropomorphic phantom work with an experimental system remained in progress.The authors note TV-based methods work well for piecewise-constant objects.Source: §IV · p. 20
• The report frames TV-superiorization as a feasibility study translating the X-ray CT approach to pCT and suggests testing other denoising cost functions.Those other functions may prove more powerful than TV, according to the authors.Source: §IV · p. 20

V. CONCLUSION

• Both TVS-DROP schemes improved spatial and density resolution while maintaining quantitative accuracy.The schemes were applied as add-ons to the standard DROP reconstruction algorithm.Source: §V · p. 21
• Removing feasibility proximity checking halved computation time and further improved image quality.The proximity-checking step was removed to reduce reconstruction time.Source: §V · p. 21
• The authors plan to implement TVS2-DROP* in future pCT image reconstruction using an experimental pCT system.Source: §V · p. 21

FIGURES

• Figure 3 compares standard DROP and two TVS reconstructions across uniform and spatial-resolution phantoms; Figure 4 plots relative error by cycle.Source: Figure 3 · p. 25, Figure 4 · p. 25
• Figures 5 and 6 compare MTF and CDF for standard DROP and reduced TVS-DROP*; TVS schemes show superior spatial resolution at any spatial frequency.Source: Figure 5 · p. 26, Figure 6 · p. 26, FIGURES · p. 26