ContiLNN: Mitigating Slice Sampling Discontinuity with Liquid Neural Networks for Medical Image Restoration
ContiLNN adds continuous-time cross-slice modeling to 2D medical image restoration backbones to handle uneven slice spacing.
The paper reports gains over Restore-RWKV on CT denoising, MRI super-resolution, and reduced-count PET restoration, with lower RMSE across all three tasks. Its Bi-CfC modules use slice-index intervals and local features to adjust propagation without numerical ODE integration. The authors also report lower latency and peak GPU memory than a similar Bi-GRU setup in matched seven-slice profiling. CT evidence is limited to one held-out patient, while PET ablations support the value of ordered propagation.
ArXiv · AI/CL/LG's note
The paper reports gains over Restore-RWKV on CT denoising, MRI super-resolution, and reduced-count PET restoration, with lower RMSE across all three tasks. Its Bi-CfC modules use slice-index intervals and local features to adjust propagation without numerical ODE integration. The authors also report lower latency and peak GPU memory than a similar Bi-GRU setup in matched seven-slice profiling. CT evidence is limited to one held-out patient, while PET ablations support the value of ordered propagation.
ArXiv · AI/CL/LG's note
score 3