VoDaSuRe: A Large-Scale Dataset Revealing Domain Shift in Volumetric Super-Resolution

1. Problem Statement

Current deep learning approaches for Volumetric Super-Resolution (SR) in medical and scientific imaging have shown impressive results, often achieving high scaling factors (e.g., 4×). However, the authors argue that this success is largely an artifact of the training methodology rather than genuine predictive capability.

• The Core Issue: Most existing SR models are trained on synthetically downsampled High-Resolution (HR) data to create Low-Resolution (LR) pairs. This creates an "over-idealistic" correspondence where the LR data is a perfect, noise-free, and artifact-free downscaled version of the HR ground truth.
• The Reality Gap: Real-world LR scans (acquired physically at lower resolution) differ fundamentally from downsampled data. They contain unique acquisition artifacts, different noise profiles, beam hardening, and signal-to-noise characteristics that synthetic degradation models fail to capture.
• The Consequence: Models trained on synthetic data fail to generalize to real LR scans. Instead of reconstructing lost microstructures, they tend to predict a "spatially averaged" or smoothed version of the input, effectively hallucinating plausible but inaccurate structures. There is a lack of large-scale, paired datasets containing physically acquired LR-HR pairs to benchmark this domain shift.

2. Methodology: The VoDaSuRe Dataset

To address the scarcity of real paired data, the authors introduce VoDaSuRe (Volumetric Dataset for Super-Resolution).

• Data Acquisition:
  • Samples: 16 diverse samples including biological materials (human femurs, vertebrae, animal bone) and non-biological materials (bamboo, cardboard, MDF, various wood species like oak, elm, larch, cypress).
  • Scanning: Samples were scanned using two different lab-CT scanners (Nikon XT H 225 and Zeiss Xradia Versa 520) tuned for specific resolution ranges.
  • Resolution Strategy: A fixed 4× resolution gap was targeted between HR and LR scans. The voxel sizes were tuned so that fine microstructures were resolved only in HR scans, while coarser structures remained visible in both, ensuring non-trivial SR tasks.
• Data Curation Pipeline:
  1. Registration: LR volumes were coarsely cropped and aligned to HR volumes using ITK-Elastix (translational and affine registration) to achieve voxel-level correspondence.
  2. Intensity Matching: To address contrast differences caused by different acquisition parameters (e.g., cone-beam dispersion), the cumulative distribution functions (CDF) of LR slices were matched to downsampled HR slices. This prevents unrealistic intensity scaling while preserving structural information.
  3. Storage: Due to the massive size of the data (~194 billion voxels total), all data was converted to the OME-Zarr format. This enables efficient out-of-core sampling of 3D patches without loading full volumes into memory.
• Dataset Scale: VoDaSuRe is the largest volumetric SR dataset to date, containing 16 paired scans (32 total volumes) with a mean volume shape of ~3330 × 1820 × 1870 voxels.

3. Experimental Setup

The authors benchmarked 8 state-of-the-art SR models (6 volumetric, 2 slice-wise) across multiple datasets:

• Datasets: CTSpine1K, LiTS, LIDC-IDRI (medical, synthetic LR), and VoDaSuRe (both synthetic downsampled and real registered LR).
• Models: RCAN, HAT, EDDSR, SuperFormer, MFER, mDCSRN, MTVNet, RRDBNet3D.
• Training: Models were trained on 2× and 4× upscaling tasks using AdamW optimizer and L1 loss.
• Evaluation Metrics: PSNR, SSIM, NRMSE, LPIPS, and Total Variation (TV) to measure high-frequency content.

4. Key Results

The experiments revealed a significant domain shift between synthetic and real data:

• Performance Drop on Real Data:
  • On medical datasets (synthetic LR), models achieved high PSNR (e.g., >40 dB at 2×).
  • On VoDaSuRe (Downsampled), performance dropped significantly (e.g., ~25 dB at 2×), indicating that even synthetic degradation on complex microstructures is harder than on smooth medical MRI/CT.
  • On VoDaSuRe (Registered/Real LR), performance dropped further (e.g., ~16–18 dB at 4×).
• Visual and Structural Differences:
  • Models trained on real LR data produced blurred, spatially averaged outputs, failing to recover fine microstructures (e.g., wood fibers, bone trabeculae).
  • Total Variation (TV) analysis confirmed a massive loss of high-frequency information in predictions trained on real data compared to HR ground truth.
  • Cross-Domain Failure: Models trained on downsampled data and tested on real LR data produced structurally plausible but inaccurate results, confirming that the "smoothness" learned from synthetic data does not transfer to real acquisition artifacts.
• Ablation Studies:
  • Misregistration: Intentional misalignment did not reproduce the smoothing effect seen in real data, proving the issue is intrinsic to the acquisition process, not just alignment errors.
  • Perceptual Loss: Adding LPIPS loss increased texture but did not resolve the domain gap or recover true microstructures.
  • Context Dependency: Using Local Attribution Mapping (LAM), the authors found that models rely more heavily on broad spatial context (higher Diffusion Index) when processing VoDaSuRe compared to medical datasets, yet this did not correlate with better accuracy.

5. Key Contributions

1. VoDaSuRe Dataset: The release of the largest paired multi-resolution volumetric dataset featuring both synthetically downsampled and physically acquired LR-HR scans across diverse structural domains (bone, wood, composites).
2. Revelation of Domain Shift: Empirical proof that current SR methods are overstated; they primarily learn to invert idealized downsampling rather than reconstructing lost physical details from real low-resolution scans.
3. Benchmarking Framework: A comprehensive benchmark showing that SOTA models fail to generalize from synthetic to real data, highlighting the need for physically grounded training data.
4. Data Pipeline: Implementation of an efficient OME-Zarr-based data loader for handling terabyte-scale 3D volumes with out-of-core sampling.

6. Significance and Conclusion

The paper fundamentally challenges the current paradigm of volumetric super-resolution research. It demonstrates that progress in deep learning-based SR is currently limited by the lack of realistic training data.

• Scientific Impact: The findings suggest that current SR models cannot reliably recover structures lost in low-resolution scans (e.g., in clinical CT or micro-CT) but instead predict smoothed averages. This has critical implications for medical diagnosis and materials science where fine structural details are crucial.
• Future Direction: The authors argue that future SR research must prioritize datasets with paired real scans (like VoDaSuRe) to develop models that can handle real-world acquisition artifacts, noise, and physics, rather than relying on synthetic degradation models.

In summary, VoDaSuRe serves as a "reality check" for the field, urging a shift from optimizing for synthetic benchmarks to solving the harder, physically grounded problem of true volumetric super-resolution.