Slice-to-Volume Reconstruction for Motion-Corrupted Medical Imaging

33 concepts31 papers3 proposals

In clinical practice, three-dimensional anatomical volumes are rarely captured all at once. Instead, modalities like Magnetic Resonance Imaging (MRI) and Optical Coherence Tomography (OCT) typically acquire data as a sequence of high-resolution 2D slices. While this sequential approach is fast, it introduces a fundamental temporal flaw: the subject often moves between the acquisition of one slice and the next.

Whether caused by a patient’s heartbeat, respiratory cycles, or the unpredictable kicks of a fetus, this inter-slice motion results in "staircase" artifacts. These misalignments disrupt the spatial continuity of the anatomy, making it difficult for clinicians to perform accurate volumetric measurements or for automated models to segment structures reliably.

The Geometry of Misalignment

The core challenge is that each 2D slice exists in its own local coordinate system, shifted or tilted by the subject's movement. As noted by Fu et al., even involuntary eye jitters in OCT can cause translational and rotational misalignments that blur clinical landmarks. In cardiac imaging, the problem is often "respiratory staircasing," where slices acquired across different breath-holds do not line up because the diaphragm returns to a slightly different position each time (Oktay et al.; Tarroni et al.).

To resolve this, researchers use Slice-to-Volume Reconstruction (SVR). The goal is to find a rigid or non-rigid transformation for every individual slice that maps it back into a consistent 3D reference frame.

Finding the Reference Frame

If the goal is to align slices, what do we align them to? Researchers have developed several strategies to establish a spatial "ground truth":

  • Geometric Anchors: In cardiac MRI, clinicians often take "long-axis" views (vertical slices) and "short-axis" views (horizontal stacks). Oktay and Stolt-Ansó et al. demonstrate that these orthogonal planes can act as geometric anchors. By maximizing the intensity agreement where these planes intersect, the entire stack can be realigned without external sensors.
  • Shape Priors: When a clear reference scan is missing, the authors of MulViMotion suggest using "shape-aware" regularization. By teaching a model what a heart should look like, the system can infer the motion correction needed to restore anatomical plausibility.
  • Self-Gating: Rather than fixing motion after the fact, Li et al. and Chen and Liu extract respiratory signals directly from the imaging data itself. This "self-gating" allows slices to be grouped by the phase of the breath, ensuring they are only compared to other slices captured in the same state.

Learning to See in 3D

Traditional registration methods, such as those using Markov Random Fields proposed by Porchetto et al., treat alignment as an optimization problem: searching for the best fit by iteratively moving slices. However, this is computationally expensive.

Recent work has shifted toward predictive registration. Hou et al. introduced SVRNet, a deep learning architecture that looks at the visual content of a single 2D slice and instantly regresses its 3D position and orientation. This is particularly vital for fetal imaging, where the subject's movement is entirely unpredictable. By breaking slices into smaller, locally rigid "patches," Kainz et al. showed that it is possible to reconstruct not just the fetal brain, but the entire moving body.

Accelerating Reconstruction

The final step of SVR is the synthesis of a high-resolution volume from the realigned, scattered slices. This is often a bottleneck in clinical workflows. To address this, Dannecker et al. proposed using 3D Gaussian Primitives. By representing the volume as a collection of differentiable Gaussians rather than a rigid grid of voxels, the authors achieved a 10x speedup in reconstruction time. This allows for high-quality fetal brain volumes to be generated in under 30 seconds, moving the technology closer to real-time diagnostic use.

By unifying alignment and reconstruction—sometimes even performing them simultaneously with segmentation, as shown by Wei et al.—these methods transform disjointed 2D captures into coherent, medically actionable 3D models.

Key Foundational Work

Respiratory Motion Correction for 2D Cine Cardiac MR Images using Probabilistic Edge Maps (2016) — contributed 2 concept elements to this direction.
Axial Alignment for Anterior Segment Swept Source Optical Coherence Tomography via Robust Low-Rank Tensor Recovery (2016) — contributed 1 concept elements to this direction.
Flexible Reconstruction and Correction of Unpredictable Motion from Stacks of 2D Images (2015) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_inter_slice_motion_artifacts_3eb5c30f,
  title = {inter-slice motion artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inter-slice-motion-artifacts/rdr_3eb5c30f691c4023af1340311cc94f63},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_inter_slice_motion_artifacts_3eb5c30f_v1,
  title = {inter-slice motion artifacts},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/inter-slice-motion-artifacts/rdr_3eb5c30f691c4023af1340311cc94f63/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake