ContiLNN: mitigating slice sampling discontinuity with liquid neural networks for medical image restoration

He et al. introduce bidirectional closed-form continuous-time modules to model irregular cross-slice medical imaging, outperforming RWKV and GRU baselines with lower latency and peak memory.

Primary source: arXiv:2610.12337: https://arxiv.org/abs/2610.12337

What’s new: Volumetric clinical scans such as CT, MRI, and PET are frequently acquired with non-isotropic voxel spacing or irregular slice intervals due to breath-hold variations, patient motion, and institutional protocols. Standard 2D restoration models treat individual slices independently, discarding cross-slice anatomical continuity, while 3D networks assume uniform spatial sampling intervals and incur high computational and memory costs. In ContiLNN: Mitigating Slice Sampling Discontinuity with Liquid Neural Networks for Medical Image Restoration (arXiv:2610.12337, 8 Oct 2026), He et al. propose ContiLNN, augmenting 2D restoration backbones with bidirectional closed-form continuous-time (Bi-CfC) liquid neural network modules.

Bi-CfC models slice propagation as a continuous dynamical system, allowing slice-index distance intervals to modulate feature gating directly without requiring numerical ODE integration. A reference-guided consistency loss aligns first-order and second-order cross-slice intensity differences to preserve delicate anatomical boundaries, while distillation from a frozen 2D backbone maintains in-plane restoration quality. Across CT denoising, MRI super-resolution, and reduced-count PET restoration, ContiLNN improves mean peak signal-to-noise ratio (PSNR) over Restore-RWKV by 0.19 dB, 1.02 dB, and 1.25 dB, respectively, with consistently lower root mean squared error. Matched seven-slice profiling shows 52.8% lower latency and 57.0% lower peak GPU memory usage compared to a bidirectional GRU baseline.

Why it matters: Clinical imaging data collected across heterogeneous scanner vendors rarely conforms to idealized isotropic grids. Adapting continuous-time dynamical models to volumetric restoration provides a principled way to ingest irregular slice intervals while keeping memory and latency low enough for practical radiological workflows.

Caveats: Evaluated on retrospective research datasets, with CT denoising results demonstrated on a single held-out patient cohort. Although mixed-gap training improves robustness on sparse or irregular slice sets, metric rankings vary across individual scanning contexts and sampling gaps.