Source-linked AI summary
SliceBridge: context-consistent repair of corrupted slice intervals in T1-weighted MRI
Jiheng Li, Michael E. Kim, Trent Schwartz, Gaurav Rudravaram, Derek B. Archer, Timothy J. Hohman, the Alzheimer's Disease Neuroimaging Initiative, Lianrui Zuo, Bennett A. Landman
TL;DR
Localized slice corruption can make an interval unreliable while surrounding MRI remains usable, creating a need for selective repair that avoids discarding or altering intact regions. SliceBridge reconstructs the interval from neighboring context with rectified flow matching and coordinated generation, improving through-plane consistency and reducing morphometric error in controlled corruptions.
Problem
Localized slice corruption can bias morphometric and anatomical analyses, while rejecting an otherwise usable scan or correcting the whole volume discards or alters unaffected regions.
Method
SliceBridge reconstructs identified corrupted axial slices from intact neighboring context using rectified flow matching, positional conditioning, and interval-correlated generation.
Results
SliceBridge reduced within-interval slice-to-slice change error by 32.9%-41.3% across interval lengths, achieved higher SSIM at every length, and reduced median regional brain-volume error from 1.95% to 1.05% in controlled corruptions.
Takeaways & Limitations
Coordinating reconstructed slices preserves axial target fidelity while improving through-plane consistency and reducing corruption-induced morphometric error.
Takeaways & Limitations
The proof-of-concept assumes corrupted intervals are already identified, reconstructs only axially, requires 16 intact context slices on both sides, and does not characterize uncertainty from multiple candidates.
Abstract
from arXiv · showhide
Structural magnetic resonance imaging (MRI) images are sometimes corrupted over a contiguous set of slices, where acquisition, motion, hardware, or reconstruction effects leave a single slice or short interval inconsistent with its neighbors while the rest of the image remains usable. Such localized corruption can bias downstream morphometric analysis, yet discarding or reacquiring an otherwise usable image is costly. We formulate this as an image restoration problem: given the location of the affected interval, reconstruct those slices from the surrounding anatomical and imaging context. We propose SliceBridge, a framework for restoring corrupted slice intervals in T1-weighted MRI using rectified flow matching conditioned on the surrounding intact slices and their relative slice positions. Through-plane consistency is encouraged by coupling the slices within the interval through interval-correlated initial noise, a shared flow time, and synchronized sampling. The restored interval is then inserted back, leaving all other slices unchanged. We trained and validated the model on 9,877 T1-weighted brain MRI volumes from four datasets and evaluated it on 581 external subjects using clean interval withholding and controlled corruptions. Compared with a matched model that reconstructed target slices independently, SliceBridge reduced error in slice-to-slice changes within repaired intervals by 32.9%-41.3% across interval lengths and achieved higher SSIM at every interval length. In controlled-corruption cases, SliceBridge reduced the median error in regional brain volume estimates produced by a downstream segmentation model from 1.95% in corrupted volumes to 1.05%.
1. INTRODUCTION
Localized slice corruption can bias downstream MRI analysis even when most of a volume remains usable. SliceBridge addresses this selective-repair problem by reconstructing corrupted axial intervals from intact neighboring context while coordinating predictions across the interval.
- Motivation: Localized corruption can alter anatomical boundaries and bias brain-volume or cortical-anatomy estimates.It may arise from acquisition, motion, hardware, or reconstruction effects and can disrupt downstream segmentation.
- Motivation: Selective restoration preserves usable slices while avoiding rejection of the entire scan or alteration of intact regions.The repair uses only intact neighboring slices and leaves all intact slices unchanged.
- Related work: Prior diffusion-MRI replacement methods rely on redundancy across measurements unavailable within a single T1-weighted volume.Structural MRI work supports same-volume contextual reconstruction but does not directly solve this setting.
- Contribution: SliceBridge reconstructs complete axial targets with rectified flow matching conditioned on intact neighboring slices and positional information.Its design targets direct axial reconstruction rather than whole-volume correction.
- Contribution: Interval-correlated noise, shared flow time, and synchronized sampling coordinate generation to improve through-plane consistency.The framework is compared with an uncoupled-flow model to isolate the contribution of interval-level coupling.
2. METHOD
SliceBridge restores identified corrupted axial intervals from intact neighboring anatomy while preserving usable slices. It combines positional conditioning with interval-level coordination so reconstructed slices remain individually plausible and coherent through-plane.
- Context and restoration: SliceBridge reconstructs corrupted axial slices directly from 16 retained context slices on each side and inserts only the generated interval back into the volume.Context candidates within another corrupted interval are excluded, so retained context need not be contiguous.
- Conditioning and reconstruction: The shared model receives intact context images, context-offset maps, target-position maps, flow time, and the current noisy target state.Positional maps encode context-to-target geometry that image channels alone do not provide.
- Design motivation: Independent slice-wise reconstruction can produce individually plausible images with abrupt intensity or anatomical changes across the repaired interval.This motivates coordinating trajectories rather than generating each target from unrelated noise and sampling paths.
- Interval coordination: SliceBridge couples target generation through interval-correlated initial noise, one shared flow time during training, and synchronized sampling during inference.The shared noise component ties starting states together while smoothed slice-specific variation preserves gradual target-to-target change.
- Training objective: For intervals of six or more slices, center-weighted supervision increases weights toward targets farther from observed context while preserving unit mean loss.Shorter intervals use ordinary mean flow-matching loss with all weights equal to one.
3. EXPERIMENTAL SETUP
The study evaluates four repair methods on externally held-out T1-weighted MRI using standardized preprocessing, clean interval withholding, controlled corruptions, and fidelity, consistency, and morphometry metrics.
- Data and preprocessing: 9,877 development volumes were retained after quality control, with 7,901 used for training and 1,976 for validation; IXI contributed 581 external evaluation volumes.
- Data and preprocessing: All volumes were reoriented, intensity-scaled to [0, 1], and represented as 256 × 256 axial slices without through-plane resampling.
- Evaluation protocols: Intervals of lengths {1, 2, 4, 6, 8, 10, 12} were withheld from clean volumes, with 16 eligible intact context slices required on each side.Intervals near sequence ends were excluded when complete bilateral context was unavailable.
- Evaluation protocols: Controlled-corruption testing applied seven corruption types to clean IXI copies while preserving the originals as references.The corruptions included dropout, intensity elevation, whiteout, noise, blur, rigid motion, and RF-like banding.
- Repair methods: Four methods were compared: linear interpolation, EG-GAN, Uncoupled flow, and SliceBridge.The Uncoupled flow versus SliceBridge comparison isolated the effect of interval-level coordination.
- Evaluation metrics: Evaluation measured reconstruction fidelity with SSIM and PSNR, through-plane consistency with internal and boundary z-gradient MAE, and downstream morphometric error with regional-volume APE.The segmentation analysis used SLANT-TICV and summarized each case by median APE across affected ROIs.
4. RESULTS
On 581 external IXI cases, SliceBridge achieved stronger reconstruction fidelity and through-plane consistency than comparison methods across interval lengths, with improved downstream morphometric accuracy in controlled corruptions. Qualitative examples showed better balance between axial appearance and coronal continuity, while the real-corruption example lacked a verified clean reference.
- Reconstruction fidelity: SliceBridge exceeded every comparator in mean foreground SSIM and PSNR at every evaluated interval length.All paired comparisons remained significant after Holm correction (all adjusted p< 0.001).
- Reconstruction fidelity: SSIM was highest near observed boundaries and lowest near the center of length-12 intervals, where SliceBridge most exceeded Uncoupled flow.The central positions were farthest from observed context.
- Through-plane consistency: 32.9%-41.3%: SliceBridge reduced mean internal z-gradient MAE relative to Uncoupled flow across intervals of length 2-12.It produced lower error in all 581 paired cases at every length, with all comparisons significant after Holm correction (all adjusted p< 0.001).
- Through-plane consistency: 3.4%-10.0%: SliceBridge reduced boundary error relative to Uncoupled flow at both interval boundaries across all lengths.Paired differences were significant at every length (all adjusted p< 0.001).
- Downstream morphometry: 0.81 percentage points: SliceBridge reduced median affected-ROI APE versus corrupted volumes, with lower error in 79.7% of paired cases.The 95% CI was 0.64-0.96, and the adjusted p-value was < 0.001; SliceBridge also significantly outperformed linear interpolation and EG-GAN.
- Downstream morphometry: The affected-ROI APE endpoint did not distinguish SliceBridge from Uncoupled flow (adjusted p= 0.370).This contrasts with the through-plane consistency and reconstruction-fidelity advantages observed for SliceBridge.
- Qualitative examples: As withheld intervals expanded from 1 to 6 and 12 slices, all methods showed increasing distortion but different failure patterns.In the example, SliceBridge preserved axial target appearance while reducing Uncoupled flow’s ripple-like through-plane discontinuities; EG-GAN retained smooth coronal consistency but lower axial fidelity.
- Qualitative examples: In the real-corruption example, post-repair SLANT-TICV labels were more continuous across the interval and more consistent with neighboring slices.Because no verified clean reference existed, this indicates improved downstream segmentation consistency rather than recovery of unknown original anatomy.
5. DISCUSSION
SliceBridge shows that repairing a corrupted slice interval requires both accurate individual-slice reconstruction and coordination across targets to preserve through-plane continuity. The method improved slice consistency and reduced corruption-induced morphometric error, while its evaluation and operating assumptions leave several practical extensions open.
- Uncoupled flow produced individually plausible target slices but substantial discontinuities within repaired intervals.
- Repairing a complete interval therefore requires reliable target reconstruction and joint coordination across multiple targets.
- Both flow models reduced morphometric error in corrupted volumes, but their regional volume-based results did not differ significantly.
- The proof-of-concept assumes corrupted intervals are already identified, reconstructs only axially, requires 16 intact context slices on each side, and does not characterize repair uncertainty.
- SliceBridge maintained high axial target fidelity, improved slice-to-slice consistency within intervals and across boundaries, and reduced corruption-induced morphometric error.