Paper review · October 12, 2026
ContiLNN: Medical image restoration that reduces slice-sampling discontinuity with Liquid Neural Networks
ContiLNN: Mitigating Slice Sampling Discontinuity with Liquid Neural Networks for Medical Image Restoration
By attaching a bidirectional CfC module that accounts for slice spacing to 2D restoration models, the authors improved CT, MRI and PET restoration quality over the baseline models.
Problem
When restoring low-dose CT, low-resolution MRI and low-count PET images, the anatomical continuity of neighboring slices is useful information. However, 2D models process each slice separately, and 3D models have a heavy memory cost. Also, when the available slices are sparse or irregularly spaced, simply counting neighboring slices does not capture the context well.
Method
The authors inserted a Bi-CfC (bidirectional closed-form continuous-time) module into some Feature paths of 2D restoration backbones (Restore-RWKV, DASMamba). The module propagates the Hidden State forward and backward along the slice order. Because it feeds the normalized gap between original slice indices into its gates, it responds to changes in sampling interval without numerical ODE integration. Training combined three losses: a reconstruction loss, a consistency loss that matches first- and second-order inter-slice intensity differences to the high-quality reference images, and a Distillation loss that makes the model follow the predictions of a frozen 2D backbone.
Results
- Averaged over 5 seeds, PSNR was higher than Restore-RWKV by 0.1907dB for CT denoising, 1.0176dB for MRI super-resolution and 1.2482dB for low-count PET restoration. RMSE was also lower on all three tasks.
- When trained on contiguous slices, it achieved higher restoration quality than a Bi-GRU with a similar parameter count and compute under every sampling condition tested.
- Measured with 7 slices, it had 52.8% lower latency and 57.0% lower peak GPU memory use than Bi-GRU.
- The authors reported that in PET ablations, the performance gain came from sequential propagation rather than simply from added capacity.
- Training with mixed gaps (mixed-gap) improved both operators in sparse and irregular contexts. However, neither was consistently better across metrics and conditions.
Limitations
The CT results are a technical result on a single evaluation patient. The CT gain (about 0.19dB) is also smaller than for MRI and PET. Under mixed-gap training, Bi-CfC was not always better than Bi-GRU. Evaluation focuses on image quality metrics such as PSNR and RMSE, and the effect on clinical reading was not examined. Backbone compatibility was tested on only two backbones.
Why it matters
Inter-slice information can be used by attaching a lightweight module to existing 2D restoration models, without a full 3D model. This could be used for post-processing restoration of low-dose or shortened scans, or of images where only some slices remain. It also has advantages in memory and latency.
Key terms
- Liquid Neural Network
- A continuous-time neural network whose rate of state change varies with the input. It is well suited to sequences with irregular observation intervals.
- CfC(Closed-form Continuous-time)
- A network that approximates Liquid dynamics with a closed-form expression. Instead of repeatedly integrating an ODE numerically, it takes the time interval as input and computes the next state directly.
- 2.5D restoration
- An approach that adds information from neighboring slices to 2D cross-section processing. It is a compromise between the efficiency of 2D models and 3D context.
- PSNR
- An image quality metric, in dB, that indicates how close a restored image is to the reference image. Higher is better.
Original
Authors: Jialei He, Enhe Liu, Sifan Song, Pengfei Jin, Jionglong Su, Hongbin Wang, Zhixiang Lu, Yanhao Huang, et al. · Posted 2026-10-08 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.