Research map / Video / Video segmentation and tracking
Data-Efficient Unsupervised Interpolation Without Any Intermediate Frame for 4D Medical Images
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes UVI-Net, an unsupervised 3D volumetric interpolation framework for 4D medical images that interpolates between two frames without intermediate ground truth, using time-domain cycle consistency and a flow-based VoxelMorph backbone with a 3D U-Net reconstruction model.
Problem
Temporal interpolation of 4D medical images (3D+time) when intermediate ground-truth frames are unavailable or costly to acquire, and when training data is scarce.
Approach
Generates virtual intermediate volumes from two real frames via flow warping, reconstructs the real start and end frames from them with a multi-scale feature-aware 3D U-Net, and uses cycle-consistency loss against the real frames as pseudo-supervision; optionally fine-tunes per test instance.
Key results
- On ACDC cardiac, UVI-Net (w/o instance-specific optimization) reaches 33.57 PSNR vs. 33.15 for the best supervised baseline MPVF and 31.56 for IDIR.
- On 4D-Lung, UVI-Net reaches 33.90 PSNR (34.00 with instance-specific optimization) vs. 32.91 for IDIR and 30.99 for SVIN.
- With a training set of only one sample, the method frequently outperforms baselines trained on the full set (over 60 samples).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add instance-specific optimization (ablation) | UVI-Net · Cardiac | PSNR: 33.57 ± 0.275 -> 33.59 ± 0.268 | Helped |
| add instance-specific optimization (ablation) | UVI-Net · Cardiac | NCC: 0.565 ± 0.007 -> 0.565 ± 0.007 | No clear effect |
| add instance-specific optimization (ablation) | UVI-Net · Cardiac | SSIM: 0.977 ± 0.001 -> 0.978 ± 0.001 | Helped |
| add instance-specific optimization (ablation) | UVI-Net · Cardiac | NMSE: 2.409 ± 0.159 -> 2.384 ± 0.157 | Helped |
| add instance-specific optimization (ablation) | UVI-Net · Cardiac | LPIPS: 1.134 ± 0.044 -> 1.066 ± 0.041 | Helped |
| add instance-specific optimization (ablation) | UVI-Net · Lung | PSNR: 33.90 ± 0.382 -> 34.00 ± 0.387 | Helped |
| add instance-specific optimization (ablation) | UVI-Net · Lung | NCC: 0.319 ± 0.003 -> 0.320 ± 0.003 | Helped |
| add instance-specific optimization (ablation) | UVI-Net · Lung | SSIM: 0.980 ± 0.002 -> 0.980 ± 0.002 | No clear effect |
| add instance-specific optimization (ablation) | UVI-Net · Lung | NMSE: 0.558 ± 0.055 -> 0.552 ± 0.055 | Helped |
| add instance-specific optimization (ablation) | UVI-Net · Lung | LPIPS: 1.512 ± 0.112 -> 1.489 ± 0.093 | Helped |
Surprising findings
- Supervised models underperform on the 4D-Lung dataset, which has far fewer intermediate frames, suggesting weak supervision quality limits supervised VFI.
Limitations
- Relies on a flow-based formulation that may still produce artifacts from warping, and virtual frames are limited to a maximum time offset of 0.5 to mitigate artifacts.
- Evaluated only on two 4D datasets (cardiac MRI and lung CT) with two-frame endpoints.
Experimental setup
Datasets: ACDC cardiac MRI 4D-Lung CT
Benchmarks: ACDC cardiac MRI 4D-Lung CT
Metrics: PSNR NCC SSIM NMSE LPIPS
Baselines: VoxelMorph TransMorph Fourier-Net+ R2Net DDM IDIR SVIN MPVF
Builds on: VoxelMorph Diffusion Deformable Model (DDM) IDIR NCC + Charbonnier losses
Closest papers by meaning
- CycleINR: Cycle Implicit Neural Representation for Arbitrary-Scale Volumetric Super-Resolution of Medical Data (CVPR 2024 · similarity 0.84)
- UniPro: Unified Multi-Mode Medical Image Segmentation from 2D Images to 3D Volumes via Propagation (NeurIPS 2026 · similarity 0.84)
- LIM: Large Interpolator Model for Dynamic Reconstruction (CVPR 2025 · similarity 0.83)
- A Unified Degradation-Robust Approach to SSL and UDA for 3D Medical Images (AAAI 2025 · similarity 0.83)
- Knowledge Transfer Scaling Laws for 3D Medical Imaging (NeurIPS 2026 · similarity 0.83)
- Learning General-purpose Biomedical Volume Representations using Randomized Synthesis (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.