Back to all papers

Bidirectional multi-scale implicit neural representation network for topology-aware lung 4D-CT registration.

November 17, 2026pubmed logopapers

Authors

Liu X,Wang Z,Chen X,Wei Y

Affiliations (2)

  • School of Control Science and Engineering, Shandong University, Jinan, Shandong Province, People's Republic of China.
  • Shandong Research Institute of Industrial Technology, Jinan, Shandong Province, People's Republic of China.

Abstract

<i>Objective.</i>Precise lung 4D-CT registration is extremely challenging due to large respiratory excursions. To capture such large-magnitude motion, existing deep learning methods typically adopt a unidirectional coarse-to-fine cascade strategy, which is vulnerable to error propagation because uncorrected coarse-level misalignments degrade subsequent finer-scale stages and often lead to spatially inconsistent deformations. While implicit neural representations (INRs) support continuous coordinate-based deformation modeling, traditional instance-specific INRs lack clinical efficiency and struggle to converge on global large deformations because of the absence of multi-scale semantic guidance. This study aims to bridge these gaps by developing a bidirectional multi-scale INR network that jointly handles large deformation modeling and localized structural fidelity.<i>Approach.</i>We propose a bidirectional multi-scale INR network. First, a generalized conditional INR framework is introduced to enable continuous deformation modeling without the need for retraining. Crucially, a bidirectional feature feedback mechanism built upon Swin Transformer injects high-level semantic priors into low-resolution streams, correcting coarse-level biases through top-down interaction. Furthermore, an anatomy-aware structural constraint is incorporated to promote anatomical consistency and local boundary alignment.<i>Main results.</i>Under the within-subject phase-held-out DIR-Lab protocol, the proposed method achieved a mean TRE of 1.28 mm and a folding rate of 0.0090%. A leave-one-case-out evaluation assessed subject-independent generalization, yielding a mean TRE of 1.53 mm and a folding rate of 0.0195%. On the independent POPI dataset, the method achieved a mean TRE of 1.84 mm and a folding rate of 0.0806%, demonstrating competitive cross-center performance. Inference took 0.33 s per case.<i>Significance.</i>The proposed framework effectively combines continuous coordinate modeling with deep learning efficiency, overcoming error propagation inherent in discrete coarse-to-fine strategies and preserving both global deformation consistency and local anatomical precision. Its sub-second inference and performance under leave-one-case-out and independent external testing support its potential for efficient image-guided radiation therapy workflows.

Topics

Four-Dimensional Computed TomographyImage Processing, Computer-AssistedLungNeural Networks, ComputerJournal Article

Ready to Sharpen Your Edge?

Subscribe to join 11k+ peers who rely on RadAISlice. Get the essential weekly briefing that empowers you to navigate the future of radiology.

We respect your privacy. Unsubscribe at any time.