Robust Multi-Scale Implicit Neural Representations for Large-Deformation Lung Registration

Johannes Bastian Gebauer, Maximilian Nielsen, Frederic Madesta, René Werner, Thilo Sentker

Keywords: Implicit Neural Representations, Deformable Image Registration, Multi-Scale Optimization, Thoracic CT

TL;DR: We enhance implicit neural representations by introducing a multi-scale dual-branch training regime to achieve state-of-the-art deformable image registration on DIR-Lab and DIR-Lab COPD datasets.

Abstract: We propose a multi-scale Implicit Neural Representation (INR) framework for dense deformable image registration, designed to stabilize convergence for large deformations while preserving precision for fine anatomical details. We model the INR as a dual-branch architecture that explicitly decomposes the motion into global and local components. The objective function is driven by mask-guided Normalized Cross-Correlation augmented by geometric and semantic regularization to ensure smooth, anatomically plausible motion. Evaluation on the DIR-Lab 4DCT thorax dataset demonstrates competitive performance with a mean Target Registration Error (TRE) below 1.0 mm. On the more challenging DIR-Lab COPDgene thorax dataset, the model achieves robust alignment with a mean TRE of 1.23 mm, yielding performance comparable to leading classical optimization frameworks. A comprehensive ablation study confirms that the dual-branch design and multi-scale optimization strategy are necessary to achieve these results, enabling stable registration with modest computational overhead. Source code is available at https://github.com/IPMI-ICNS-UKE/DUAL-INR-DIR.

Primary Subject Area: Image Registration

Secondary Subject Area: Unsupervised Learning and Representation Learning

Registration Requirement: Yes

Reproducibility: https://github.com/IPMI-ICNS-UKE/DUAL-INR-DIR

Visa & Travel: No

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Replace NNN with your OpenReview submission ID., Includes \documentclass{midl}, \jmlryear{2026}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package., Did not use the times package., Use the correct spelling and format, avoid Unicode characters, and use LaTeX equivalents instead., Any math in the title and abstract must be enclosed within $...$., Did not override the bibliography style defined in midl.cls and did not use \begin{thebibliography} directly to insert references., Avoid using \scalebox; use \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., No separate supplementary PDF uploads., Acknowledgements, references, and appendix must start after the main content.

Submission Number: 283