ContiLNN: Mitigating Slice Sampling Discontinuity with Liquid Neural Networks for Medical Image Restoration
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: Today's paper: "ContiLNN: Mitigating Slice Sampling Discontinuity with Liquid Neural Networks for Medical Image Restoration".
Jane: The gist The ContiLNN model augments two-dimensional medical image restoration with bidirectional liquid dynamics over ordered slice positions to mitigate slice sampling discontinuity and improve fidelity across CT, MRI,
Tom: First, who's behind it and why it matters.
Paper summary: Tom: So, to recap this paper, "ContiLNN: Mitigating Slice Sampling Discontinuity with Liquid Neural Networks for Medical Image Restoration," the authors are proposing a new way to fix image restoration problems caused by inconsistent slice sampling in medical scans.
Jane: They argue that anatomical continuity provides complementary information for medical image restoration, but using it requires accounting for local anatomy and how different slices are sampled. So they introduce ContiLNN, which adds bidirectional closed-form continuous-time modules to two dimensional restoration backbones for cross-slice modeling while keeping the in plane feature extraction.
Lu: The thesis is that slice-index intervals can modulate gates determined by local features and hidden states, allowing the propagation to respond to sampling variations without needing numerical ODE integration. They are using bidirectional liquid neural dynamics along the slice axis, denoted as e zero through e six and hidden states h zero through h six.
Meng: And they use these bidirectional liquid neural dynamics for a forward state update and a backward state update, which is what creates the ordered window positions t(C) that links the anatomical ordering to the dynamics used to refine features.
Tom: They also incorporate reference-guided consistency, which aligns first- and second-order cross-slice intensity differences with high quality references to preserve anatomical variation, alongside distillation from a frozen two dimensional backbone to keep in plane fidelity.
Jane: The goal is restoration of an ordered slice sequence X = x zero x K-one where the ContiLNN 2D backbone and bidirectional LNN dynamics e t and per-slice features h t restore the sequence = zero K-one.
Lu: And they use the direction specific closed form continuous time update, which propagates states h t using features e t and normalized intervals delta t derived from gaps g t.
Meng: It’s about using these mechanisms to handle the missing ordered context effectively, which is what makes this work for scenarios like independent slice sampling.
Tom: This approach addresses the issue of omitting ordered context in scenarios where slices are independently sampled, and it shows how bidirectional dynamics can propagate that necessary anatomical information across those gaps.
Jane: It’s about making sure that even when you only have a subset of the full ordered sequence, ContiLNN uses all the available information to give a good restoration result.
Lu: It’s a novel way to integrate continuous time modeling into image restoration architectures specifically tailored for handling anatomical continuity in sequential data.
Meng: So this paper is showing how you can use liquid neural dynamics and closed-form updates to solve the slice sampling discontinuity problem in medical imaging.
Tom: And it sets up a comparison against other state space methods like MambaIR or Restore-RWKV, showing where this specific mechanism shines in terms of fidelity and efficiency.
Jane: It's a strong piece of work because it directly tackles the challenge of using anatomical continuity robustly in sequence restoration tasks.
Conclusion: Tom: So, wrapping up on "ContiLNN: Mitigating Slice Sampling Discontinuity with Liquid Neural Networks for Medical Image Restoration," the authors are essentially showing how to use liquid neural networks to model context across slices to fix problems caused by inconsistent slice sampling in medical imaging.
Jane: They argue that anatomical continuity is useful, but exploiting it requires accounting for local anatomy and sampling variations. ContiLNN uses bidirectional closed-form continuous-time modules and slice-index intervals to propagate context through the sequence.
Lu: The main implication is that this suggests a path forward for using liquid dynamics as a practical approach to incorporating anatomical continuity into two dimensional restoration networks in these settings.
Meng: For me, it means we should look at how this structure can be adapted for more general sequence modeling problems where you have sequential data with missing or inconsistent order.
Tom: I think the authors are demonstrating that when you train ContiLNN under contiguous conditions, it achieves higher fidelity than a bidirectional gated recurrent unit with similar parameter counts and arithmetic costs across all tested sampling conditions.
Jane: So, in simple terms, this work shows that modeling the discontinuity through these dynamics can give us better results for CT, MRI, and PET modalities when dealing with varying slice sampling.
Lu: It validates the idea of using liquid neural dynamics as a practical tool for incorporating anatomical continuity into 2D restoration networks <ref:2610.12337#pg2>.
Meng: From an engineering side, it’s about showing that we can achieve better performance with lower latency and memory use compared to some existing recurrent models when things line up nicely.
Tom: And the paper also points out that while they tested index-defined sampling, a limitation is that physical slice spacing metadata were unavailable in their preprocessed data.
Jane: That's a fair point, so the method is very powerful for handling ordered inputs, but its real strength lies in how it manages the discontinuity itself.
Jialei He, Enhe Liu, Sifan Songb, Pengfei Jinc, Jionglong Sub, Hongbin Wanga, Zhixiang Lub, Yanhao Huanga, Anteng Caia, Zhengyong Jiangb, Jiaman Dinga*, S. Kevin Zhoud*, Jinfeng Wanga*
Faculty of Information Engineering and Automation, Kunming University of Science and Technology · Academy of Artificial Intelligence and Advanced Technology, Xi’an Jiaotong-Liverpool University · Department of Radiology, Massachusetts General Hospital and Harvard Medical School · School of BME & Suzhou Institute for Advanced Research, Center for Medical Imaging, Robotics, Analytic Computing & Learning (MIRACLE), University of Science and Technology of China
cs.CV, cs.AI
Submitted: 2026-10-08
Updated: 2026-10-08
Code: https://github.com/Yaziwel/Restore-R
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 88/100
The gist: The gist The ContiLNN model augments two-dimensional medical image restoration with bidirectional liquid dynamics over ordered slice positions to mitigate slice sampling discontinuity and improve
Key concepts
- Bidirectional Closed-Form Continuous-Time (Bi-CfC) Modules
- These modules use bidirectional liquid neural dynamics to model the relationship between adjacent slices. They allow information to flow both forward and backward across the slice order, creating a continuous context that helps smooth out inconsistencies when slices are sampled irregularly.
- Slice-Index Intervals
- These intervals modulate gates within the network based on local features and hidden states. This mechanism allows the model to adapt its propagation strategy dynamically to how slices are actually sampled, enabling it to respond effectively to sampling variations without needing complex numerical integration.
- Cross-Slice Consistency
- This regularization technique aligns the first and second-order intensity differences between adjacent slices with high-quality reference data. This ensures that the anatomical variations preserved across different slices remain consistent and anatomically plausible in the final restored image.
Terminology
Summary
The gist The ContiLNN model augments two-dimensional medical image restoration with bidirectional liquid dynamics over ordered slice positions to mitigate slice sampling discontinuity and improve fidelity across CT, MRI, and PET modalities
How it works
ContiLNN introduces bidirectional closed-form continuous-time (Bi-CfC) modules that model cross-slice context through liquid neural dynamics while retaining in-plane feature extraction Slice-index intervals modulate gates determined by local features and hidden states, allowing propagation to respond to sampling variations and anatomical context This closed-form update avoids iterative numerical integration, which is particularly relevant when the separation between observations varies The forward and backward states communicate information between preceding and following observations, linking the anatomical ordering of the input to the dynamics used to refine its features
The model incorporates several complementary regularizers alongside reconstruction supervision to preserve anatomical variation These include cross-slice consistency, which aligns first- and second-order intensity differences with high-quality references, and distillation from a frozen two-dimensional backbone to retain in-plane fidelity The training objective combines reconstruction loss, cross-slice consistency loss, and 2D backbone distillation loss
Key Contributions
The authors detail three main contributions to the field of medical image restoration
-
Liquid dynamics for medical image restoration, where ContiLNN propagates anatomical information across ordered slices through bidirectional liquid dynamics and slice-index intervals condition the closed-form updates
-
Augmentation of 2D restoration backbones, where Bi-CfC modules refine selected feature pathways while retaining the backbone’s in-plane blocks, supported by cross-slice intensity consistency and 2D backbone distillation
-
Evaluation across modalities and sampling settings, assessing performance on CT, MRI, and PET under various conditions including sparse and irregular context
Experimental Results
Experiments across CT denoising, MRI super-resolution, and reduced-count PET restoration demonstrate improved fidelity compared to the two-dimensional baseline ContiLNN improves mean peak signal-to-noise ratio (PSNR) by 0.1907 dB for CT, 1.0176 dB for MRI, and 1.2482 dB for PET The improvements persist across five training seeds, with gains greater for MRI and PET
Under contiguous training, Bi-CfC achieves higher fidelity than a bidirectional gated recurrent unit (Bi-GRU) with similar parameter counts and arithmetic costs across all tested sampling conditions, while exhibiting lower measured latency and memory use The mixed-gap study shows that broader sampling exposure benefits both operators, with closely matched performance under that training strategy
Qualitative and Robustness Analysis
Qualitative results show that ContiLNN preserves local structure across the three modalities, retaining clearer soft-tissue interfaces in CT and following MRI contours more closely in HQ references Case-level analyses confirm that gains extend across held-out cohorts, with median PSNR gains of 1.0443 dB for MRI and 1.1711 dB for PET Furthermore, context perturbations show larger losses when distinct local neighbors are removed or replaced than when their order is reversed
The model's robustness to sampling variations is confirmed in the mixed-gap study, where Bi-CfC gains 0.5525 dB in PSNR at g = 2 and 0.4032 dB under irregular sampling The PET ablations support ordered propagation beyond additional pointwise capacity
Efficiency and Limitations
Bi-CfC reduces median window latency from 380.92 ms to 179.85 ms (a 52.8% reduction) compared to Bi-GRU, despite closely matched parameter counts and arithmetic costs The limited-patient PET experiments show that budget-matched training improves PSNR by 0.5914 dB at the 5% patient fraction However, physical slice-spacing metadata are unavailable in the preprocessed data, meaning experiments assess index-defined sampling rather than adaptation to calibrated anatomical distances
In conclusion, ContiLNN augments 2D medical image restoration with bidirectional liquid dynamics over ordered slice positions to mitigate slice sampling discontinuity and improve fidelity across CT, MRI, and PET modalities The findings support Bi-CfC as the preferred sequence operator in the evaluated ContiLNN framework and liquid dynamics as a practical approach to incorporating anatomical continuity into 2D restoration networks.
Improvements for AI systems
- Bold header: Slice-axis context modeling via Bi-CfC modules
The improved system can model cross-slice context through bidirectional closed-form continuous-time (Bi-CfC) modules to model cross-slice context through liquid neural dynamics while retaining in-plane feature extraction.
This allows the network to respond dynamically to sampling variations and anatomical context
by conditioning updates on the normalized slice-index interval.
- Bold header: Enhanced fidelity via dual regularization
The system can preserve both local anatomy and in-plane structure by employing two complementary regularizers: cross-slice consistency aligns first- and second-order intensity differences with high-quality references to preserve anatomical variation,
and distillation from a frozen two-dimensional backbone helps retain in-plane fidelity.
- Bold header: Robustness to sparse/irregular sampling
The system can maintain high performance under challenging data conditions because the model uses slice-index intervals
to condition its updates, enabling it to handle both sparse and irregular context using mixed-gap training,
which the paper shows improves restoration quality over standard contiguous training.
- Bold header: Efficient computational scaling for sequence propagation
The system can achieve a balance between performance and resource usage because the Bi-CfC operator is shown to have lower measured latency and memory use
compared to alternatives like Bi-GRU, even when parameter counts are similar, allowing for practical implementation in resource-constrained environments.
- Bold header: Data efficiency through budget matching
The system can be effectively trained with limited paired data because the approach uses budget-matched training,
which improves PSNR by achieving 0.5914 and 0.5889 dB at the 5% and 10% patient fractions, respectively.
Related papers
- Loss Knows Best: Detecting Annotation Errors in Videos via Loss Trajectories
- AnchorWeave: World-Consistent Video Generation with Retrieved Local Spatial Memories
- Benchmarking the Robustness of Foundation Models for Mammography under Domain Shift
- MambaX-Net: Dual-Input Mamba-Enhanced Cross-Attention Network for Longitudinal MRI Segmentation
- TeleOCR: Navigating Document Parsing Across Digital and Camera-Captured Documents
- A Survey on Efficient Vision-Language-Action Models