PaperlayerView exact revision on arXiv

Source-linked AI summary

MRI-to-CT Translation of Spinal Images via Anatomy-Guided Pseudo-Pairing and Adversarial Diffusion

Eduardo Nieto, Ibai Genua, Davide Scorza, Karen López-Linares

arXiv:2610.37134v1cs.CVcs.AI

TL;DR

MR provides soft-tissue information without ionizing radiation, while CT better depicts bone, but paired data and 3D slice consistency remain challenges for MR-to-CT translation. The paper combines vertebral matching, adversarial diffusion, and volumetric refinement; matching improves vertebral metrics, while refinement improves coherence with some realism trade-offs.

  • Problem

    MR-to-CT translation is constrained by scarce paired datasets and 2D methods’ limited inter-slice consistency and 3D anatomical coherence.

    Source: §1 · p. 1

  • Method

    The framework matches vertebral anatomy across unpaired scans to create pseudo-pairs for SynDiff training, then applies 3D GAN post-processing to improve adjacent-slice consistency.

    Source: §1 · p. 2, §3.2.1 · p. 5

  • Results

    Vertebral matching improved vertebra-wise FID and mean binary and multilabel 3D Dice; post-processing raised TotalSegmentator 3D DSC from 0.564 to 0.575 but worsened FID realism metrics.

    Source: §4 · p. 9

  • Takeaways & Limitations

    Volumetric refinement reduced discontinuities and produced smoother, more anatomically coherent vertebral reconstructions, while matching better delineated spinous processes.

    Source: §4 · p. 9

    Sacrum and posterior vertebral elements remain failure-prone, thoracic performance is less stable, and post-processing mildly smooths non-bone tissues, reducing global full-image realism.

    Source: §4 · p. 11

Abstract

from arXiv · show

Magnetic Resonance Imaging (MRI) is routinely used to diagnose spinal disorders, but surgical planning often requires Computed Tomography (CT) for better visualization of vertebral bone anatomy. Synthesizing CT from MRI could avoid additional radiation exposure and streamline preoperative assessment. We propose a three-stage framework for spine MRI-to-CT synthesis. First, an anatomy-guided vertebral matching system identifies similar sagittal MRI and CT slices using spine segmentations and aligns them through rigid registration, creating pseudo-paired data from unpaired scans. Second, these pairs train an adversarial diffusion model for MRI-to-CT translation. Third, synthesized slices are assembled into 3D volumes and refined by a network that improves inter-slice continuity and structural coherence. The framework was trained on over 10,000 unpaired sagittal MRI and CT slices and evaluated on 1,911 slices from 83 MRI volumes spanning all spinal regions. Assessment combined distribution-based measures with vertebral segmentation metrics. Compared with diffusion alone, the proposed approach increased binary 3D vertebral Dice from 0.666 to 0.739. Volumetric post-processing reduced visible discontinuities between stacked slices and improved vertebral coherence in axial and coronal views. Combining anatomy-guided pseudo-pairing, adversarial diffusion and 3D refinement improves spine MRI-to-CT synthesis. Vertebral matching primarily enhanced anatomical alignment and fidelity, while volumetric refinement improved 3D consistency. These findings support further development of MRI-derived CT-like visualization for preoperative planning, particularly to preserve fine intensity details while maintaining volumetric coherence.

1. Introduction

MR and CT provide complementary spinal information, but their selective use and CT’s radiation exposure motivate CT-like synthesis from MR. Existing methods face limited paired data and slice-wise 2D consistency, motivating a three-stage framework for unpaired spine translation.

1. Introduction

  • MR depicts soft tissues without ionizing radiation, whereas CT provides detailed bone anatomy but carries radiation exposure, motivating CT-like information from MR.Together, MR and CT support vertebral delineation, neural-structure localization, and vertebral-morphology assessment in surgical planning.Source: §1 · p. 1
  • Existing methods often use 2D slices, limiting inter-slice consistency and 3D coherence, while paired datasets are scarce and acquisition differences can compromise alignment.The text identifies respiratory motion, tissue deformation, and patient positioning as potential sources of misalignment.Source: §1 · p. 1
  • The proposed framework creates pseudo-pairs through vertebral matching, trains SynDiff across spinal regions, and applies 3D GAN post-processing to improve consistency across adjacent slices.The paper describes the pseudo-pairs as generated from different patients and the training data as unpaired.Source: §1 · p. 2

2. State of the art

The review traces medical image synthesis from GANs toward Transformer and diffusion approaches, then identifies unresolved full-spine correspondence and 3D consistency challenges. These gaps motivate the paper’s vertebral matching and volumetric refinement steps.

2. State of the art

  • The review describes GANs, Transformers, and diffusion models as successive methodological families for medical image synthesis, with diffusion and Transformer approaches capturing long-range dependencies and local detail.It also notes spine-specific GAN and diffusion studies, including lumbar MR-to-CT translation and SynDiff’s ability to leverage unpaired data.Source: §2 · p. 2, §2.2 · p. 2, §2.3 · p. 3

2.1. GAN-based approaches

  • GAN methods evolved from paired-data approaches to unpaired CycleGANs, whose cycle-consistency loss constrains cross-modal mappings while enabling training on unpaired MR and CT data.The review also notes remaining training instability and residual inconsistencies that can compromise robustness and structural precision.Source: §2.1 · p. 2

2.2. Transformers and diffusion models

  • Transformer- and diffusion-based methods are described as offering better synthesis quality than earlier GAN approaches, especially in anatomical coherence and image fidelity.The review also notes higher computational complexity and memory requirements as practical constraints for high-resolution imaging.Source: §2.2 · p. 2

2.3. MR-to-CT translation of spine images

  • Full-spine imaging still faces unresolved vertebral correspondence across unpaired datasets and slice-to-slice consistency in 3D volumes.These challenges informed the use of SynDiff as a baseline and the addition of vertebral matching and 3D consistency enhancement.Source: §2.3 · p. 3

3. Materials and methods

The paper describes its materials and methods as the data and pipeline for unpaired MR-to-CT spine image translation.

3. Materials and methods

  • The methods section introduces the data and proposed pipeline for unpaired MR-to-CT spine image translation.Source: §3 · p. 3

3.1. Dataset collection and preparation

The study assembles publicly available scans spanning cervical, thoracic, and lumbar regions, then filters and preprocesses them into sagittal slices for model development. Its preparation pipeline uses segmentation-guided subvolume extraction and patient-level dataset splits.

3.1. Dataset collection and preparation

  • Publicly available datasets were selected to cover cervical, thoracic, and lumbar spine regions, with sagittal T1-weighted MR acquisitions preferred for global assessment and vertebral visualization.Source: §3.1 · p. 3
  • The MR collection combined regional datasets and yielded 560 scans, while CT datasets provided analogous regional coverage from multiple sources.Source: §3.1 · p. 3
  • Scans with fewer than five vertebrae or severe acquisition artifacts were excluded, leaving a cohort of 558 MR and 1136 CT scans from 10 open-access datasets.Source: §3.1 · p. 3

3.1.1. Preprocessing

  • Scans were reoriented, resampled to 1 mm x 1 mm x 2.5 mm, and intensity-normalized; CT was clipped to [-1000, 1000] and MR underwent bias field correction.Source: §3.1.1 · p. 3
  • Segmentation-guided cropping extracted 288x288xD subvolumes to cover up to five vertebrae, with an additional inferior-origin crop used to complete spinal-column coverage.Source: §3.1.1 · pp. 3–4
  • Redundancy controls limited overlap between extracted subvolumes and restricted MR subvolume segmentation similarity to 90%, while CT subvolumes were not similarly restricted.Source: §3.1.1 · p. 4
  • The 2D model used sagittal slices from the subvolumes and discarded slices containing fewer than four vertebrae.Source: §3.1.1 · p. 4
  • Patient-level splitting targeted 70% training, 20% validation, and 10% test sets while preserving representative distributions of vertebral-region combinations.Source: §3.1.1 · p. 4

3.2. Proposed approach

The framework generates pseudo-paired MR–CT slices through anatomy-guided vertebral matching, translates MR slices with SynDiff, and uses the resulting pairs to train the translation model.

3.2. Proposed approach

  • The three stages are vertebral matching of unpaired sagittal scans, SynDiff-based 2D translation, and 3D post-processing for consistency across adjacent slices.The post-processing network is a 3D GAN with a ResUNet generator and PatchGAN discriminator.Source: §3.2 · p. 4

3.2.1. Vertebral matching system

  • A 1x25 vertebral-proportion descriptor filters candidate MR–CT slices before rigid registration and weighted DSC comparison select the best match.The search begins with the 50 most similar candidates, limited to slices differing by at most one vertebra; the selected CT slice is removed from later searches, and the elbow analysis identifies 50 as the efficiency–accuracy trade-off.Source: §3.2.1 · p. 5
  • When the first 50 candidates contain no match because the same vertebrae are unavailable, the system expands the pool and permits a greater vertebral difference.Matching can include redundant CT patches, while MR patches are filtered to reduce similar spine information.Source: §3.2.1 · p. 5, Figure 2 · p. 5

3.2.2. MR to CT translation using SynDiff

  • SynDiff combines a non-diffusive approximation module with a diffusive denoising module, with both modules structured as CycleGANs.The study compared region-matched slices without vertebral matching against pseudo-pairs created using vertebral structural similarity.Source: §3.2.2 · p. 6
  • Both SynDiff setups used 256x256 padded sagittal inputs and were trained until spine-related metrics stabilized.Training used batch size 1, learning rate 1 × 10−4, two residual blocks in the non-diffusive generator, 1000 direct diffusion steps, and four denoising steps.Source: §3.2.2 · p. 6

3.3. 3D contextual consistency enhancement

A supervised 3D post-processing network refines stacked synthetic CT volumes to reduce inter-slice discontinuities while preserving in-slice image quality.

3.3. 3D contextual consistency enhancement

  • The refinement network targets local discontinuities that can arise when independently translated 2D sagittal slices are reassembled into a 3D volume.It is intended to improve anatomical continuity across adjacent vertebral structures without changing the core translation architecture.Source: §3.3 · p. 6
  • Training used 670 real CT subvolumes and pseudo-CT inputs created by matching each slice to anatomically similar slices from other CT scans, disrupting inter-slice continuity.The original coherent CT subvolumes served as supervised targets.Source: §3.3 · p. 6
  • Testing used 83 synthetic CT volumes assembled from translated MR sagittal slices in the test split.The reported network training used batch size 1, with generator and discriminator learning rates of 4 × 10−4 and 8 × 10−5, respectively.Source: §3.3 · pp. 6, 8
  • A residual encoder–decoder generator and volumetric PatchGAN discriminator form the supervised 3D GAN used for refinement.The generator’s skip connections preserve anatomical detail, while the discriminator evaluates overlapping volumetric patches.Source: §3.3 · p. 6, Figure 5 · p. 8
  • The generator objective combines adversarial, L1, SSIM, perceptual, and inter-slice consistency losses.Interslice consistency is the L1 distance between consecutive sagittal slices; perceptual loss uses features from a pretrained 3D ResNet-18.Source: §3.3 · p. 7

3.4. Validation methods

Validation combines distribution-level image comparisons, vertebral segmentation measures, and vertebra-specific intensity-distribution comparisons.

3.4. Validation methods

  • Evaluation combines probability-distribution, task-specific, and atlas-based metrics on full images and specific vertebral regions.The measures address image similarity as well as anatomical utility and fidelity.Source: §3.4 · p. 8

3.4.1. Distribution-based metrics

  • Raw MMD values have no fixed scale and are not directly comparable across different distributions, so the study reports relative MMD against the real MR–CT discrepancy.The relative measure is the synthetic CT–CT MMD divided by the real MR–CT MMD.Source: §3.4.1 · p. 8

3.4.2. Task-specific metrics

  • Vertebral utility is assessed with DSC between segmentations from the input MR and corresponding synthetic CT, assuming spatial correspondence is preserved.TotalSegmentator supports volumetric and slice-level evaluation and indirectly measures 3D coherence through spatial context.Source: §3.4.2 · p. 8

3.4.3. Atlas-based metrics

  • Atlas-based evaluation compares synthetic vertebral HU distributions with reference distributions extracted from real CT vertebral masks.Partially visible border vertebrae are excluded; EMD measures global intensity shifts and JSD measures histogram-shape similarity.Source: §3.4.3 · p. 9

4. Results and discussion

The vertebral matching system improves quantitative results over the baseline, while 3D post-processing improves volumetric coherence but smooths generated volumes and worsens realism metrics.

4. Results and discussion

  • Vertebral matching improves results over baseline, with lower vertebra-wise FID and higher mean binary and multilabel 3D Dice scores.The authors attribute this improvement to anatomically coherent pseudo-pairs providing stronger supervision than the unpaired setting.Source: §4 · p. 9
  • The matching system reduces rMMD from 0.135 to 0.053 versus baseline, while Full FID increases from 77.84 to 80.88 and global image quality remains comparable.Reconstructed volumes also benefit in 3D Dice score, with more anatomically coherent volumes than the baseline.Source: §4 · p. 9
  • 0.575 versus 0.564 in TotalSegmentator 3D DSC after refinement, alongside increased full-image FID (80.88 to 87.41) and vertebral FID (33.64 to 50.37).The post-processing network visually reduces discontinuities in axial and coronal planes, but smooths vertebrae and surrounding tissues.Source: §4 · p. 9
  • Refinement partially improves cervical fidelity despite greater variability, improves upper thoracic EMD and JSD, but does not reduce lower thoracic errors and worsens sacral values.Lumbar values remain similar before and after processing, with the best EMD and JSD across the spine, though L1 degrades slightly; T7 is highlighted as challenging.Source: §4 · p. 10
  • 3D refinement reduces slice-stacking discontinuities and improves vertebral correspondence with MR segmentations across anatomical planes.The figure compares MR ground-truth segmentations, synthetic CT segmentations before refinement, and refined results.Source: §4 · p. 9, Figure 9 · p. 11
  • The sacrum and posterior vertebral elements remain failure-prone, thoracic performance is less stable, and smoothing harms non-bone realism; absent paired ground truth also limits evaluation.The authors caution that FID's natural-image feature extractor may not fully capture medical-image anatomy.Source: §4 · p. 11

5. Conclusions

The pipeline's vertebral matching improved alignment and structural fidelity in synthetic CT, while 3D post-processing removed slice discontinuities but reduced sagittal intensity accuracy and realism. The findings establish a baseline for more coherent synthetic imaging and motivate refinement that preserves fine detail.

5. Conclusions

  • The fully automatic pipeline matches and aligns unpaired sagittal slices through spine segmentations, trains SynDiff, then post-processes predicted volumes for spatial coherence.The approach establishes a baseline for further development of spatially coherent synthetic imaging.Source: §5 · pp. 11–12
  • Vertebral matching improved cross-modal alignment and structural fidelity, particularly in vertebral regions, despite slightly worse overall image-quality metrics.The study prioritizes anatomical accuracy in key regions for preoperative planning over translation of other structures.Source: §5 · p. 11
  • The 2D approach showed discontinuity artifacts across slices and struggled to maintain anatomical coherence in axial and coronal views.Validated volumetric segmentation achieved acceptable 3D segmentation, but qualitative results still showed the artifacts.Source: §5 · p. 12
  • 3D post-processing eliminated slice artifacts and slightly improved vertebral fidelity across planes, but reduced sagittal vertebral intensity accuracy and overall realism.The intensity and realism costs limit the refinement's preservation of 2D image quality.Source: §5 · p. 12
  • Future refinement should reduce discontinuities while preserving fine-grained detail and 2D intensity fidelity; fully 3D or hybrid 2D/3D approaches are proposed directions.The paper also identifies anatomical priors and multimodal constraints as potential directions.Source: §5 · p. 12
  • Synthetic CT could potentially reduce additional CT acquisitions in preoperative planning, though clinical adoption faces regulatory and validation challenges.CT-like information derived from preoperative MRI is discussed in relation to image-guided procedures.Source: §5 · p. 12
Loading 2610.37134v1…