Research map / 3D vision / 3D generation
LiFT: Lifted Inter-slice Feature Trajectories for 3D Image Generation from 2D Generators
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 LiFT, which lifts a 2D slice synthesizer into a 3D volume generator via a learned depth-indexed trajectory module: tri-planar drifting loss for unconditional MR generation (LiFT-U) and a BiGRU z-context mixer with supervised losses for paired MR-to-CT and missing-MR translation (LiFT-C).
Problem
High-resolution 3D medical image synthesis is expensive for fully volumetric generators, while slice-wise 2D generators produce anatomically inconsistent volumes across depth.
Approach
Keep the synthesizer predominantly 2D and add a small learned through-plane module: a depth mapper trained with tri-planar distributional drifting loss (unconditional), or a bidirectional GRU over slice bottleneck descriptors trained with paired reconstruction and z-derivative losses (paired).
Key results
- Unconditional BraTS 128^3: LiFT-U reaches FID x1e3 of 0.066 vs 0.154 for WDM, with inference memory 0.41 GB vs 2.55 GB.
- Missing-MR BraTS: LiFT-C inference 1.16 s per volume vs 156.4 s for cWDM (~135x), with average PSNR 28.31 dB vs published cWDM 28.42 dB.
- MR-to-CT SynthRAD: LiFT-C reaches 57.50 HU MAE vs 59.40 for no-mapper ablation, and full-volume delta-z MAE drops from 42.73 to 38.41.
Surprising findings
- The LiFT-C trajectory module changes through-plane coherence more than voxel-wise MAE on MR-to-CT (MAE gain small).
- Missing-MR comparison against cWDM uses published aggregate numbers, so no formal paired test against cWDM is possible.
Limitations
- Performance is coupled to the quality of the underlying 2D slice generator or translator; in-plane errors (bone/air) remain large.
- Evaluation limited to brain imaging with no multi-institutional or domain-shift testing.
- LiFT-U guarantees are distributional; the cWDM comparison relies on published aggregates without paired statistics.
Experimental setup
Datasets: BraTS 2023 GLI SynthRAD2023 Task 1 Brain
Benchmarks: BraTS 2023 GLI SynthRAD2023
Metrics: FID MS-SSIM inference memory PSNR SSIM MAE NCC delta-z MAE delta-z correlation inference time
Baselines: 2.5D LDM 3D DDPM 3D LDM HA-GAN WDM cWDM pix2pix CBAM3D-UNet Pix2pix-ResNet
Builds on: drifting models WDM pix2pix Fourier positional encoding
Closest papers by meaning
- Better Tokens for Better 3D: Advancing Vision-Language Modeling in 3D Medical Imaging (NeurIPS 2025 · similarity 0.86)
- Fully Convolutional Slice-to-Volume Reconstruction for Single-Stack MRI (CVPR 2024 · similarity 0.85)
- CycleINR: Cycle Implicit Neural Representation for Arbitrary-Scale Volumetric Super-Resolution of Medical Data (CVPR 2024 · similarity 0.84)
- PolyPose: Deformable 2D/3D Registration via Polyrigid Transformations (NeurIPS 2025 · similarity 0.84)
- TCAM-Diff: Triplane-Aware Cross-Attention Medical Diffusion Model (AAAI 2025 · similarity 0.83)
- Intraoperative 2D/3D Image Registration via Differentiable X-ray Rendering (CVPR 2024 · 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.