Biophysics-Enhanced Neural Representations for Patient-Specific Respiratory Motion Modeling DOI: 10.59275/j.melba.2026-1ba1 Volume: 2026
\firstname Jan \surname Boysen \aff 1 \orcid 0009-0005-1308-6696, \firstname Hristina \surname Uzunova \aff 3 \orcid 0000-0001-8854-3517, \firstname Heinz \surname Handels \aff 1,2 \orcid 0000-0002-3499-4328, \firstname Jan \surname Ehrhardt \aff 1,2 \orcid 0009-0004-6804-5587 · 2026
Read Biophysics-Enhanced Neural Representations for Patient-Specific Respiratory Motion Modeling DOI: 10.59275/j.melba.2026-1ba1 Volume: 2026 by \firstname Jan \surname Boysen \aff 1 \orcid 0009-0005-1308-6696, \firstname Hristina \surname Uzunova \aff 3 \orcid 0000-0001-8854-3517, \firstname Heinz \surname Handels \aff 1,2 \orcid 0000-0002-3499-4328, \firstname Jan \surname Ehrhardt \aff 1,2 \orcid 0009-0004-6804-5587 (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-22123.txt.
Introduction
Boysen and co-authors model how a patient's breathing moves tissue during radiotherapy. Precise dose placement matters, and lung motion adds real uncertainty. Their model, PRISM-RM, uses implicit neural representations as a motion surrogate. It needs no fixed reference breathing state. The motion field is continuous in space and time and stays diffeomorphic. Biophysical constraints keep the estimated motion plausible beyond the training data. The authors report interpolation on par with their earlier model and better extrapolation. Sequential registration methods still extrapolate better, they note.
Catalog record
| CREATOR | \firstname Jan \surname Boysen \aff 1 \orcid 0009-0005-1308-6696, \firstname Hristina \surname Uzunova \aff 3 \orcid 0000-0001-8854-3517, \firstname Heinz \surname Handels \aff 1,2 \orcid 0000-0002-3499-4328, \firstname Jan \surname Ehrhardt \aff 1,2 \orcid 0009-0004-6804-5587 |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43850 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.22123
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Biophysics-Enhanced Neural Representations for Patient-Specific Respiratory Motion Modeling DOI: 10.59275/j.melba.2026-1ba1 Volume: 2026 \firstname Jan \surname Boysen \aff 1 \orcid 0009-0005-1308-6696, \firstname Hristina \surname Uzunova \aff 3 \orcid 0000-0001-8854-3517, \firstname Heinz \surname Handels \aff 1,2 \orcid 0000-0002-3499-4328, \firstname Jan \surname Ehrhardt \aff 1,2 \orcid 0009-0004-6804-5587 Abstract A precise spatial delivery of the radiation dose is crucial for the treatment success in radiotherapy. In the lung and upper abdominal region, respiratory motion introduces significant treatment uncertainties, requiring special motion management techniques. To address this, respiratory motion models are commonly used to infer the patient-specific respiratory motion and target the dose more efficiently. In this work, we investigate the possibility of using implicit neural representations (INR) for surrogate-based motion modeling. Therefore, we propose physics-regularized implicit surrogate-based modeling for respiratory motion (PRISM-RM). Our new integrated respiratory motion model is free of a fixed reference breathing state. Unlike conventional pairwise registration techniques, our approach provides a trajectory-aware spatio-temporally continuous and diffeomorphic motion representation, improving generalization to extrapolation scenarios. We introduce biophysical constraints, ensuring physiologically plausible motion estimation across time beyond the training data. Our results show that our trajectory-aware approach performs on par in interpolation and improves the extrapolation ability compared to our initially proposed INR-based approach. Compared to sequential registration-based approaches both our approaches perform equally well in interpolation, but underperform in extrapolation scenarios. However, the methodical features of INRs make them particularly effective for respiratory motion modeling, and with their performance steadily improving, they demonstrate strong potential for advancing this field. \melbaid 2026:008 \melbaauthors Boysen et al. \firstpageno 148 \melbayear 2026 \datesubmitted 2025-07 \datepublished 2026-03 \melbaspecialissue MELBA–BVM 2025 Special Issue \melbaspecialissueeditors Andreas Maier, Thomas Deserno, Heinz Handels, Klaus Maier-Hein, Christoph Palm, Thomas Tolxdorff, Katharina Breininger \ShortHeadings Biophysics-Enhanced Respiratory Motion ModelingBoysen, Uzunova, Handels, Ehrhardt \affiliations \num 1 \addr German Research Center for Artificial Intelligence (DFKI), Lübeck, DE \num 2 \addr Institute of Medical Informatics, University of Lübeck, Lübeck, DE \num 3 \addr Clinic for Orthopedics and Orthopedic Surgery, University Medicine Greifswald, Greifswald, DE keywords Physics-Informed Machine Learning, Implicit Neural Representations, Registration, Respiratory Motion Modeling 1 Introduction \enluminure The aim of radiotherapy is to accurately deliver a high dose of radiation to the tumor while sparing as much healthy tissue as possible. In tumor treatment in the lung and upper abdomen, respiratory motion is one of the main sources of error, and respiratory motion management such as respiratory gating or tracking is required to compensate for tumor motion during treatment ( 11 ) . The ideal scenario for managing respiratory motion is to directly observe and track the tumor and organs at risk (OARs) during radiation therapy. However, achieving this requires suitable imaging technologies integrated into the linear accelerator that either provide functional information, such as PET ( 15 ) , or high soft tissue contrast, such as MRI ( 6 ) . The clinical application of techniques like MRI-guided radiation therapy (MRIgRT) is currently limited by the need for additional infrastructure, increased treatment and personnel costs, and the lack of medium- to long-term clinical outcome data ( 18 ) . Less complex and more cost-effective linac-integrated imaging techniques such as MV/kV imaging enable sufficiently reliable and robust marker-less tracking only for specific cases and tumor locations ( 34 ) . An alternative to direct imaging and tracking of tumor and OARs is indirect monitoring using internal or external surrogate signals. Examples of internal surrogate signals include implanted fiducials tracked by kV imaging or electromagnetic transponders ( 8 ; 4 ) , while external surrogate signals include spirometers, pressure belts, or optical tracking systems. In this work, we rely on external surrogate signals, although the proposed methodology is equally applicable to internal surrogate signals. Then, based on a surrogate signal that can be acquired during treatment, deformable motion models allow for the estimation of respiratory motion in an entire anatomical region, including the tumor and nearby OAR ( 19 ) . Such models are usually generated based on previously acquired 4D image data using a sequential process including two-stages: first, the motion is estimated from the imaging data using deformable image registration and second, a correspondence model approximates the relationship between the surrogate signal and the motion using regression modeling ( 19 ; 27 ) . Later work ( 20 ; 17 ) proposed an integrated model that jointly optimizes image registration and correspondence model fitting, improving the reconstruction of 4D image data. The majority of approaches assume simple linear or polynomial relationships between the surrogate signals and local deformations. The estimation of motion from 4D image data is usually done by pairwise registration of the dynamic 3D image frames to a fixed image, acquired at a reference breathing state ( 12 ) . A variety of optimiza-tion- and learning-based registration approaches has been applied specifically to thorax images, see 30 for an overview. Recently, implicit neural representations (INRs) have gained attention in the field of medical image registration ( 25 ) , performing on par with the state-of-the-art for deformable lung registration ( 29 ; 36 ) . INRs offer a continuous representation of deformation, and their capacity for analytical differentiation provides significant benefits for implementing PDE-based regularizers and loss functions ( 7 ) . Closely related to physics-informed neural networks (PINNs) ( 21 ) , biophysical constraints can be introduced into the registration model as proposed by 1 and, moreover, INRs naturally fit into a diffeomorphic registration setting ( 24 ; 26 ; 16 ; 33 ; 25 ) . Another line of research uses INRs for motion-compensated image reconstruction and benefits in particular from the implicit continuity regularization of INRs and the ability to reconstruct from sparsely and irregularly sampled data ( 14 ; 22 ; 35 ) . Although some of these approaches include the cardiac or respiratory phase as a temporal dimension, the motion is only implicitly modeled. Despite the success of the approaches in the context of image registration and motion-compensated image reconstruction, there is rarely any work utilizing INRs for surrogate-based motion modeling. In this work, we investigate implicit neural representations for creating patient-specific motion models that represent the lung deformation over an entire breathing cycle. Specifically, based on a 4D image dataset with associated surrogate signals, we aim to learn a model that is able to estimate the internal motion state from unseen surrogate signals. This work builds upon our previous conference paper ( 5 ) , in which we developed an integrated INR-based motion model that uses spatial position and surrogate signals as input dimensions to predict local motion. The network is capable of learning complex functions to model the relationship between surrogate signals and internal breathing motion. It learns the respiration-induced motion deformations from sparsely sampled image data by registering images from different breathing states to a reference state. This paper extends our previous work in several aspects: First, a much more detailed mathematical derivation and description of the approach is presented in the context of surrogate based motion modeling. More importantly, based on the transport equation, we extend our INR-based approach to a diffeomorphic, time-continuous respiratory motion model that no longer requires a fixed reference image for pair-wise registration. Instead, we introduce a group-wise diffeomorphic INR-based registration, enabling motion estimation between arbitrary respiratory phases.