Source-linked AI summary
Differentiable Jitter Correction using Deep Learning-based Image Quality Metric for Phase-Contrast Micro-CT
Junan Chen, Yiting Jia, Joscha Maier, Dominik John, Sami Wirtensohn, Imke Greving, Silja Flenner, Matthias Wieczorek, Julia Herzen
TL;DR
Jitter undermines the high-resolution structural information provided by phase-contrast µCT, while reference-based quality metrics are unavailable at deployment. The paper introduces differentiable parallel-beam geometry optimization with a CNN-predicted VIF metric and background-restricted TV regularization, recovering lost structural detail across varied specimens and acquisition settings.
Problem
Phase-contrast µCT requires jitter correction, but reference-based image-quality metrics need motion-free volumes unavailable during deployment.
Method
The method combines analytic parallel-beam reconstruction gradients, a compact CNN predicting VIF from one degraded volume, and background-restricted TV regularization.
Results
Across four cases spanning in-distribution organs, an unseen specimen, and a real acquisition, the pipeline recovers structural detail lost to jitter.
Takeaways & Limitations
Reference-free differentiable quality optimization is a viable route to jitter correction across varying morphologies and imaging conditions.
Takeaways & Limitations
The evaluation relies on injected jitter whose assumed motion model matches the one used to generate the data, leaving validation against measured jitter for future study.
Abstract
from arXiv · showhide
This paper proposes a fully differentiable jitter correction method for X-ray phase-contrast micro computed tomography using a deep learning-based image quality metric that estimates and compensates per-projection rigid jitter directly from the acquired projection data, without a pre-scan motion-free reference. The approach builds on a gradient-based auto-focus strategy adapted to parallel-beam geometry. A set of candidate objective functions is benchmarked in a controlled study, and the sensitivity of the visual information fidelity (VIF) metric to the jitter artifact is verified with the target phase-contrast data. To operate without a clean reference, a compact 3D convolutional neural network is trained to predict the VIF score from a single corrupted volume. A spatially selective total variation penalty applied exclusively to the image background is introduced to penalize spurious high-frequency structures that otherwise emerge during optimization. Experiments on biological specimens acquired at different synchrotron beamlines are conducted. Evaluation uses jitter motion applied to simulated and experimentally acquired projection data. The result confirms that the integrated pipeline reliably recovers fine structural detail lost due to jitter, with generalization demonstrated across morphologically distinct samples.
1 Introduction
Phase-contrast µCT provides highly sensitive micrometer-scale soft-tissue imaging but imposes tight geometric constraints. This work addresses jitter correction with differentiable, reference-free quality optimization adapted to parallel-beam acquisitions.
- Phase-contrast µCT uses refraction and interference to provide superior contrast for weakly absorbing biological structures at micrometer-scale resolution.
- The method targets jitter correction in parallel-beam phase-contrast µCT, where projection geometry and specimen diversity differ substantially from clinical CT.
- Differentiable gradient-based optimization derives the parallel-beam reconstruction gradient and estimates geometry directly from acquired projection data.
- A compact 3D CNN predicts VIF from a single degraded volume, removing the need for a paired clean reference during inference.
- A background-restricted TV penalty suppresses spurious high-frequency textures while avoiding smoothing genuine internal sample structure.
2 Methods
The method differentiates the reconstruction with respect to per-projection geometry, allowing gradient-based optimization of a quality-driven jitter-correction objective. It combines a learned reference-free quality metric with normalized, background-selective TV regularization to suppress artifacts while preserving specimen structure.
- Differentiable CT geometry optimization: The cost combines image quality from the reconstructed volume with a weighted regularization term, C(P) = Q(µ(P)) + βR(P).Q evaluates reconstruction quality, while R counteracts degeneracies of the primary quality criterion.
- Differentiable CT geometry optimization: Gradient-based optimization updates per-projection geometry using an analytic reconstruction gradient for parallel-beam CT.The reconstruction is modeled as bilinear projection sampling, enabling derivatives with respect to projection matrices and a 3 × 4 geometry gradient for each projection.
- Differentiable CT geometry optimization: TV regularization is normalized by the initial reconstruction, producing a dimensionless term whose weight can transfer across volumes of different sizes.The current volume’s TV is divided by the initial uncorrected volume’s TV before applying β.
- Differentiable CT geometry optimization: A spatially selective TV penalty acts only outside the specimen, suppressing non-physical background textures while leaving internal structure untouched.The background region is defined by a binary mask Ωbg, and its sub-gradient is backpropagated through the same Jacobian as the quality term.
- Deep learning-based reconstruction quality metric: DL-VIF predicts reference-based VIF from a single corrupted volume, making the quality term usable without a motion-free reference at inference time.VIF is selected over SSIM because it is described as more structure-agnostic and information-focused for these artifacts.
- Metrics evaluation: Jitter experiments perturb projection matrices independently by random rotations and translations, then reconstruct the corrupted projections using the same operator.The severity parameters θmax and smax control rotational and translational jitter magnitude.
3 Results
The results compare image-quality metrics for jitter correction, assess reference-free DL-VIF prediction on held-out specimens, and evaluate optimization across simulated and experimentally acquired data. VIF is sensitive near low jitter, DL-VIF tracks ground-truth VIF across anatomies, and the proposed optimization improves over corrupted inputs without ground-truth access.
- 3.1 Metrics evaluation with controlled jitter levels: VIF drops steeply at mild jitter before saturating, whereas SSIM decreases more uniformly across the six simulated motion levels.The contrasting responses reflect VIF’s sensitivity to early high-frequency information loss and SSIM’s emphasis on global structural correlation.
- 3.1 Metrics evaluation with controlled jitter levels: VIF is steepest near the optimization-relevant high-quality regime, while saturation occurs after high-frequency content is already destroyed.Its natural-scene-statistics formulation also supports transfer across unseen morphologies despite the nonlinear response.
- 3.2 DL-VIF training result: Predicted DL-VIF scores correlate strongly with ground-truth VIF across brain, heart, worm, and testis specimens, with r = 0.965, 0.990, 0.975, and 0.973, respectively.Figure 3 evaluates held-out specimens across varying translation, rotation, and motion-duration severity.
- 3.2 DL-VIF training result: The heart predictions show a roughly constant downward bias despite the highest correlation, indicating strong association but weaker agreement with the identity line under distribution shift.The paper attributes the offset to differences between test-specimen appearance statistics and the training distribution.
- 3.3 Optimization result: Reference-based SSIM and VIF achieve the highest SSIM scores, while DL-VIF consistently improves over the motion-corrupted baseline without ground-truth access at inference.For the heart, the reference-based scores are 0.9108/0.9079; for the worm, they are 0.9731/0.9799.
- 3.3 Optimization result: Adding TV regularization improves sharpness-based correction for every specimen and further closes the gap between DL-VIF and reference-based upper-bound objectives.All modes use per-projection six-parameter jitter optimization with Adam; background TV is used in the regularized modes.
4 Discussion and conclusion
The proposed differentiable pipeline enables reference-free jitter correction in parallel-beam phase-contrast µCT and generalizes across specimens and acquisition conditions. Its performance is demonstrated alongside explicit limitations involving surrogate optimization, training resolution, motion-model assumptions, and other geometry errors.
- Discussion and conclusion: The pipeline combines analytic parallel-projection differentiation, a compact single-volume CNN for VIF prediction, and background-restricted normalized TV regularization.The CNN achieved PLCC > 0.96 on held-out specimens, while the selective regularizer constrains featureless regions without smoothing genuine internal structure.
- Discussion and conclusion: SSIM increased from 0.42–0.57 uncorrected to 0.73–0.94 after correction across four test cases, including an out-of-distribution specimen and a real worm acquisition.For the real worm acquisition, SSIM was 0.9411 after correction versus 0.5740 before correction.
- Discussion and conclusion: Transferability is addressed through specimens spanning five morphological classes and an architecture that compresses feature maps to 2×2×2 for global degradation prediction.This removes the need for fixed-anatomy shape priors and reduces the first fully connected layer by roughly four orders of magnitude.
- Discussion and conclusion: Optimizing the learned quality term can generate unseen texture in empty regions, because the surrogate is optimized outside its fitted distribution.Background-only TV regularization is reported as an effective remedy, whereas global TV can remove genuine internal structure.
- Discussion and conclusion: The evaluation remains bounded by 128^3-voxel training-label preprocessing, injected jitter matching the assumed motion model, and untested drift or calibration errors.The paper proposes higher-resolution training, validation with measured jitter and independent tracking, and benchmarking against synchrotron projection-alignment methods.
5 Acknowlegdement
The authors acknowledge the facilities, beamtime, dataset contributions, and technical support that enabled the research.
- Acknowlegdement: The study used experimental facilities from Helmholtz-Zentrum Hereon and DESY, with part of the dataset collected at PETRA III.Beamtime was allocated under proposals I-20221419 and I-20240219.