LiFT: Efficient Factorized Volume Synthesis with Lifted Inter-Slice Feature Trajectories

TL;DR

Factorized synthesis: LiFT keeps the synthesis pathway two-dimensional and learns how slice representations evolve along depth with a small inter-slice trajectory module.
Unconditional (LiFT-U): a depth mapper turns a global latent code into per-slice codes for a frozen 2D generator, trained with a tri-planar drifting loss that matches axial, coronal, and sagittal feature distributions.
Paired translation (LiFT-C): a bidirectional GRU runs along depth at every bottleneck location of a 2D U-Net and is trained with direct supervision from the target volume.
Efficient: 390× faster than 3D MedDiffusion for unconditional brain MR, and 133× faster than cWDM for missing-MR synthesis at matched quality (within 0.008 dB PSNR).
Coherent: inter-slice context lowers through-plane error by 10–13% in MR-to-CT synthesis and reduces the axial-bias SSIM gap by 48% in missing-MR synthesis.

Abstract

High-resolution 3D medical image generation remains challenging because fully volumetric models are computationally expensive, while efficient 2D slice generators often fail to preserve anatomical consistency across the third dimension. We propose LiFT, a framework for Lifted inter-slice Feature Trajectories that factorizes 3D volume synthesis into per-slice image generation and inter-slice trajectory learning. Rather than using a fully volumetric synthesis backbone, LiFT treats a volume as an ordered trajectory in feature space, capturing how anatomical structures appear, transform, and disappear across depth. We demonstrate the LiFT framework in both unconditional and conditional settings. In unconditional generation, a tri-planar drifting loss matches feature distributions between generated and real volumes across three orthogonal planes; in paired translation, a bidirectional z-context mixer trained with paired target supervision supplies through-plane coherence while preserving per-slice fidelity. We evaluate LiFT on BraTS 2023 for unconditional and missing-modality MR synthesis, and on SynthRAD2023 for MR-to-CT translation. LiFT achieves competitive image quality while retaining the efficiency of predominantly 2D synthesis, operating up to 390× faster than a volumetric diffusion baseline and consistently improving through-plane coherence through explicit inter-slice modeling. These results establish lightweight inter-slice trajectory learning as an efficient approach to high-resolution 3D medical synthesis.

Method

A 2D synthesis model whose slice representations pass through the LiFT inter-slice trajectory module before decoding into the generated volume. — LiFT factorizes volume synthesis into a 2D synthesizer and an inter-slice trajectory module.

Unconditional

LiFT-U: distributional lifting

A depth mapper maps a global latent ξ and a Fourier depth encoding γ(d) to a conditioning code c_d for each slice of a frozen 2D generator. The tri-planar drifting loss pulls features of generated axial, coronal, and sagittal slices toward the real-volume feature bank, so no paired target is needed.

Paired translation

LiFT-C: supervised lifting

A bidirectional GRU runs along depth at every location of a 2D encoder–decoder’s spatial bottleneck and returns its context through a zero-initialized residual. Encoder, decoder, and skip connections stay two-dimensional; training uses the registered target volume.

Results

Unconditional T1n generation (BraTS 2023 GLI)

Axial, coronal, and sagittal slices across depth from a LiFT-U volume and a WDM volume. — Axial, coronal, and sagittal reformats across depth from LiFT-U and WDM.

FID from 2048-d MedicalNet ResNet-50 features, 1,000 generated vs. 1,000 real volumes. WDM FID is the published value.
Method	FID ×10³ ↓	Params (M) ↓	Time / vol. (s) ↓
WDM	0.154	57.9	26.92
MAISI-v2	0.294	201.4	0.506
3D MedDiffusion	0.043	166.3	15.37
LiFT-U	0.066	30.3	0.039

Missing-modality MR synthesis (BraTS 2023 GLI)

T2-FLAIR synthesis by pix2pix, context-disabled LiFT-C, and LiFT-C in three planes, with absolute error maps. — T2-FLAIR synthesis in three planes with absolute error maps. Context disabled is the same LiFT-C checkpoint with the context module turned off.

Averages over the four target contrasts on the 219-subject validation split.
Method	Avg. PSNR ↑	Avg. SSIM ↑	Params (M) ↓	Time / vol. (s) ↓
pix2pix	26.351	0.9345	54.4	0.053
Fast-cWDM	28.233	0.9456	81.5	15.54
cWDM	28.425	0.9525	81.5	154.8
LiFT-C	28.417	0.9528	41.3	1.166

MR-to-CT synthesis (SynthRAD2023 Task 1 Brain)

Through-plane coherence on the 36-subject evaluation split; Δz is the adjacent-slice difference along depth (HU).
Method	Δz MAE full ↓	Δz MAE bone ↓	Δz MAE air ↓	Δz corr. ↑
Pix2pix-UNet	52.20	180.23	199.71	0.575
Pix2pix-ResNet	44.91	159.54	186.91	0.650
CBAM3D-UNet	39.88	141.10	142.37	0.713
Context disabled	42.73	151.73	172.21	0.675
LiFT-C	38.369	134.968	149.722	0.729

Citation

@inproceedings{zhang2026lift,
  title     = {Efficient Factorized Volume Synthesis with Lifted Inter-Slice Feature Trajectories},
  author    = {Zhang, Xinhe and Zhang, Yuyang and Jin, Pengfei and Marin-Llobet, Arnau and Li, Na and Li, Quanzheng},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}