LiFT generates 3D medical images 135x faster using 2D slices

A team led by Xinhe Zhang has published LiFT (Lifted inter-slice Feature Trajectories), a novel framework that enables high-resolution 3D medical image generation without the computational burden of full volumetric models. Instead of modeling the entire 3D distribution end-to-end, LiFT factorizes the process into per-slice 2D image generation and a lightweight inter-slice trajectory learner. This treats a 3D volume as an ordered path in feature space, capturing how anatomical structures appear, transform, and disappear across depth. A tri-planar drifting loss aligns generated trajectories with real ones for unconditional generation, while a bidirectional z-context mixer provides through-plane coherence for paired translation tasks.

LiFT was evaluated on BraTS 2023 (unconditional and missing-modality MR) and SynthRAD2023 (MR-to-CT). Results show it preserves per-slice quality and approaches the reconstruction quality of the state-of-the-art cWDM model at roughly 135× lower inference cost. It also markedly improves through-plane coherence compared to a no-mapper ablation in MR-to-CT translation. This demonstrates that lightweight inter-slice trajectory learning is a viable and efficient path to high-resolution 3D medical synthesis, potentially enabling faster, cheaper medical imaging workflows.