PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep preferential evolution guided by pairwise comparisons enables motion correction in photoacoustic tomography

Karteekeya Sastry, Yousuf Aborahama, Junhao Zhu, Manxiu Cui, Lihong V. Wang

arXiv:2610.12273v1physics.med-pheess.SP

TL;DR

In 3D PAT, extended mechanical scanning and low-quality sub-reconstructions make motion correction difficult. DPE uses a learned image-quality evaluator to guide evolutionary search, and its comparator generalizes more robustly from simulation to in-vivo images than an absolute scorer while mitigating natural-motion artifacts in palm images.

  • Problem

    Extended acquisition makes 3D PAT vulnerable to motion, while low-quality sub-reconstructions make direct motion tracking challenging.

    Source: Schematic of DPE-based motion correction · p. 3

  • Method

    DPE couples evolutionary search with a learned image-quality evaluator to optimize intra-acquisition motion correction in 3D PAT.

    Source: Discussion · p. 12

  • Results

    A same-target comparator yields robust simulation-to-experiment generalization compared with an absolute scorer, and comparator-based DPE mitigates artifacts from natural unconstrained motion in in-vivo palm images.

    Source: Discussion · p. 12

  • Takeaways & Limitations

    Comparator-based DPE mitigated natural-motion artifacts in human palm images in a setting motivated by experimental training-data scarcity in PAT.

    Source: Discussion · p. 12

    Methods may fail to recover structures heavily obscured by artifacts.

    Source: Discussion · p. 13

Abstract

from arXiv · show

Motion artifacts in three-dimensional photoacoustic tomography (PAT), caused by extended mechanical scanning of sparse arrays, degrade image quality. Since sparse sampling results in low-quality sub-reconstructions, tracking motion directly is challenging. While final image quality could in principle guide artifact correction, conventional regularizers do not accurately reflect photoacoustic image quality. Here, we introduce deep preferential evolution (DPE), a derivative-free optimization framework where a learned image comparator guides evolutionary search over reconstruction parameters. By learning relative quality differences between same-target image pairs rather than absolute quality scores, the comparator generalized more robustly from simulation to in-vivo images (after unlabeled domain calibration) than an absolute scorer. Comparator-based DPE successfully corrected synthetic motion across diverse anatomies and backgrounds, and it mitigated artifacts from natural, unconstrained motion in human palm images. These results demonstrate that pairwise learned comparison can provide a transferable optimization objective for high-dimensional image restoration when labeled experimental training data are scarce.

Results

The results examine comparator-based DPE as a way to correct motion using final reconstructed image quality, including its generalization and correction effectiveness across anatomical targets and acquisition conditions. They cover synthetic motion across in-vivo anatomies and background levels, as well as natural, unconstrained motion in human palms.

Schematic of DPE-based motion correction

  • DPE uses final-image quality to estimate a corrective motion trajectory because sparse, low-quality sub-reconstructions make direct motion tracking challenging.The estimated trajectory enables image-domain correction of sub-reconstructions to mitigate artifacts and restore image quality.Source: Schematic of DPE-based motion correction · pp. 3–4, Figure 1 · p. 4

Comparison of DPE paradigms

  • Across all five reported metrics—PCC, SSIM, PSNR, CNR, and RMS voxel deviation—comparator-corrected images outperformed scorer-corrected images.The comparison covered in-vivo test baselines.Source: Comparison of DPE paradigms · p. 6, Figure 2 · p. 7
  • Simulation accuracies were 91.3±1.9% for the scorer and 89.3±1.8% for the comparator; in-vivo, scorer accuracy fell 15.8 percentage points to 75.5±3.3%, versus comparator’s 7.4-point drop to 81.9±2.0%.Source: Comparison of DPE paradigms · p. 6
  • Comparator-corrected in-vivo palm images were sharper and had fewer artifacts than scorer-corrected images, with higher SSIM and PCC relative to baseline.Both paradigms recovered key features, but the comparator produced the better correction.Source: Comparison of DPE paradigms · p. 6, Figure 2 · p. 7

Effectiveness of DPE-based motion correction across anatomical targets and background levels

  • DPE effectiveness was demonstrated across palm, wrist/forearm, and breast images, including acquisitions at 1064 and 755 nm with varying backgrounds and vascular contrasts.Source: Effectiveness of DPE-based motion… · p. 8
  • Repeated correction of two synthetic-motion cases with distinct initialization seeds demonstrated stability of the stochastic optimization process.The repeated runs are shown in Supplementary Fig. 5.Source: Effectiveness of DPE-based motion… · p. 8
  • DPE recovered image quality across six cases spanning palm, wrist/forearm, and breast anatomies, two illumination wavelengths, and varying background levels despite training only on simulated images without substantial background.The synthetic motion trajectories were smooth and temporally correlated, an out-of-distribution motion model.Source: Effectiveness of DPE-based motion… · pp. 8, 10, Figure 3 · p. 9

DPE-based motion correction of natural unconstrained human motion

Discussion

DPE couples evolutionary search with a learned image evaluator for intra-acquisition motion correction in 3D PAT. Its comparator generalizes more robustly than an absolute scorer and supports correction across varied conditions, including natural motion in palm images.

Discussion

  • DPE couples evolutionary search with a learned image-quality evaluator for intra-acquisition motion correction in 3D PAT.The framework uses a comparator trained to assess relative structural differences between images sharing an underlying target.Source: Discussion · p. 12
  • The comparator generalized more robustly than an absolute scorer and supported correction over approximately 300 motion parameters across varied targets, backgrounds, and motion models, plus natural-motion palm images.The authors report successful mitigation of artifacts in in-vivo palm images with natural, unconstrained motion.Source: Discussion · p. 12
  • DPE may extend to other PAT optimization tasks, including aberration correction and speed-of-sound tuning, and to modalities such as X-ray CT and MRI.The learned comparator was also applied zero-shot to automated speed-of-sound tuning.Source: Discussion · p. 12
  • Motion correction may relax the cost, acquisition-time, and image-quality trade-off, supporting parsimonious arrays with less compromise to image quality.In the experimental system, fewer arcs lower cost but require a larger rotation angle and longer acquisition time.Source: Discussion · p. 12
  • DPE uses the neural network as an evaluator rather than a generator, while physical reconstruction confines candidates to plausible solutions and limits hallucination risk.The discussion contrasts this setup with direct image-domain correction approaches.Source: Discussion · pp. 12–13
  • A single in-vivo image currently takes approximately one hour to correct over 200 iterations, motivating faster optimization.The authors suggest global optimization followed by a faster local method as one possible approach.Source: Discussion · p. 13
  • The current implementation assumes rigid motion; soft-deformation models may be essential for targets such as the breast affected by heartbeat and breathing.Keeping the search-space dimension manageable while exploring deformable models is identified as a future step.Source: Discussion · p. 13
  • Comparator performance may improve through optimizer-in-the-loop fine-tuning on adaptively sampled candidates in addition to randomly corrupted training images.The proposed training extension would use candidates sampled by the optimizer.Source: Discussion · p. 13

Methods

The methods cover 3D PAT acquisition, synthetic motion-corrupted data, and controlled in-vivo motion experiments, alongside derivative-free motion-trajectory search guided by learned image-quality scoring or comparison. They also describe image standardization, evaluation metrics, and participant cohorts.

Experimental data acquisition

  • Experimental acquisition used a rotating-arc hemispherical array (26 cm, 2.25 MHz) and 1064 nm Nd:YAG or 755 nm Alexandrite illumination; dual-wavelength pulses were interleaved.Source: Experimental data acquisition · pp. 13–14

Synthetic data generation

  • Synthetic training data used 500 VascuSynth vessels and 100 rigid-motion corruptions per vessel, with translation and rotation increments sampled within 0.3 mm and 0.3° spheres.Source: Synthetic data generation · p. 14
  • Motion extent in each synthetic corrupted image was scored as mean temporal coordinate variance across voxels, corresponding to the squared RMS voxel deviation.Source: Synthetic data generation · p. 14

Network architecture and training

  • Both networks fused spatial convolutions and Fourier features; the comparator learned pairwise labels and used swapped image order with inverted targets to encourage antisymmetry.The spatial branch captured local geometry and vessel continuity, while Fourier features represented non-local phase shifts and global blur.Source: Network architecture and training · p. 15

In-vivo image reconstruction

  • For in-vivo demonstrations, baseline images were treated as motion-free before synthetic motion was added, using random trajectories or smooth Gaussian-filtered random walks capped at 0.225 mm and 0.225° per pulse.Source: In-vivo image reconstruction · p. 16
  • Unlabeled self-standardization of in-vivo volumes incurred only a minor cost in comparison accuracy, and the synthetic-motion application discrepancy was negligible in this setting.The discrepancy was between applying synthetic motion to reconstructed image stacks rather than underlying targets.Source: In-vivo image reconstruction · pp. 16–17

Deep preferential evolution (DPE) framework

  • DPE estimates a corrective trajectory by optimizing 50 time steps of 3D translations and rotations, with the first step fixed, leaving 294 parameters.The method treats reconstruction as derivative-free optimization over motion parameters.Source: Deep preferential evolution (DPE)… · p. 17
  • CMA-ES searches trajectories within bounds based on maximum inter-step motion, implicitly constraining smoothness; optimization uses 200 iterations and a population of 22.The same bounds were applied to natural-motion demonstrations, based on observed motion severity.Source: Deep preferential evolution (DPE)… · p. 17
  • Scorer-based DPE ranks candidate images by predicted scores, whereas comparator-based DPE uses pairwise tournament outcomes and a wins-minus-losses fitness normalized by comparisons.Comparator outcomes are binarized before candidate fitness is computed.Source: Deep preferential evolution (DPE)… · p. 17

Image quality metric evaluation

  • Reconstruction quality was evaluated using PCC, SSIM, PSNR, CNR, and RMS voxel deviation; image metrics were computed after registering volumes to baseline images.PSNR and SSIM used the baseline image's 99.9th-percentile voxel amplitude as dynamic range.Source: Image quality metric evaluation · p. 18

Imaging protocols

  • The study included nine adults; the in-vivo test set comprised four subjects, natural-motion images came from two, calibration statistics from another two, and breast images from an independent volunteer.No participants were excluded from analysis.Source: Imaging protocols · p. 18

Data availability

The paper identifies which acquired data, supporting materials, and motion-correction resources will be publicly available, while noting that reconstruction code is proprietary.

Data availability

Code availability

  • DPE-based correction code for acquired images in Figs. 4a–4d and trained comparator weights will be made publicly available upon publication.Source: Code availability · p. 24
  • Reconstruction code is not publicly available because it is proprietary and used in licensed technologies.Source: Code availability · p. 24

Contributions

The authors describe the study design, framework development, experiments, analysis, and manuscript work, and disclose one author's financial interests in three companies that did not support the work.

Contributions

  • K.S., Y.A., and L.V.W. designed the study, while K.S., Y.A., and M.C. designed and developed DPE.Source: Contributions · p. 24
  • J.Z. and K.S. performed experiments; K.S. and Y.A. analyzed data and wrote the manuscript with input from all authors, and L.V.W. supervised and revised it.Source: Contributions · p. 24

Competing interests

  • L.V.W. reported financial interests in Microphotoacoustics Inc., CalPACT LLC, and Union Photoacoustic Technologies Ltd., which did not support the work.Source: Competing interests · p. 25
Loading 2610.12273v1…