Paper deep dive
Personalising Deformable Image Registration with Clinical Context and Deformation History for Adaptive Radiotherapy
Caiwen Jiang, Songlin Zhao, Yuzhen Ding, Jia Mi, Samir H. Patel, Alexandra Bennett, Keldon K. Lin, Terence T. Sio, Jonathan B. Ashman, Lisa A. McGee, Jean-Claude M. Rwigema, William G. Rule, Sameer R. Keole, Sujay A. Vora, William W. Wong, Nathan Y. Yu, Michele Y. Halyard, Steven E. Schild, Safia K. Ahmed, Daniel J. Ma, Robert L. Foote, Michelle A. Neben Wittich, Thomas J. Whitaker, Haibo Lin, Dinggang Shen, Wei Liu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/15/2026, 4:35:27 AM
Summary
The paper introduces CAP-DIR, a clinical-context-aware and patient-adaptive deformable image registration method for adaptive radiotherapy. CAP-DIR integrates source-side contours, dose distributions, clinical text, and longitudinal deformation history to improve registration accuracy. Evaluated on 16,361 CT examinations from 3,473 patients, CAP-DIR significantly outperformed existing learning-based methods (like CorrMLP) in organ-at-risk (OAR) alignment, contour propagation usability, and dose-mapping accuracy.
Entities (8)
Relation Signals (6)
CAP-DIR → improves → OAR alignment
confidence 95% · CAP-DIR increased internal-test organ-at-risk-region normalized cross-correlation from 84.76% to 96.92%... CAP-DIR also retained a low OAR folding ratio...
CAP-DIR → outperforms → CorrMLP
confidence 95% · CAP-DIR increased internal-test organ-at-risk-region normalized cross-correlation from 84.76% to 96.92% relative to the strongest learning-based comparator [CorrMLP]...
CAP-DIR → uses → clinical context
confidence 95% · CAP-DIR, integrating source contours, dose, clinical text and prior-registration features.
CAP-DIR → uses → deformation history
confidence 95% · CAP-DIR... integrating... prior-registration features... longitudinal memory adapts the inference to the anatomical trajectory already observed...
CAP-DIR → evaluatedon → Mayo Clinic
confidence 90% · Model development used patient-level partitions of the Mayo Clinic Arizona and Minnesota cohorts.
CAP-DIR → reduces → dose-mapping error
confidence 90% · In a controlled benchmark, CAP-DIR reduced global dose-mapping error by 34.5% relative to the strongest comparator.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Adaptive radiotherapy must account for anatomical changes during treatment, but doing so depends on reliably mapping targets, organs at risk and dose across serial images. Existing registration methods treat each image pair independently, without using radiotherapy-specific context or the patient's preceding anatomical trajectory. We developed Clinical-context-Aware and Patient-adaptive Deformable Image Registration (CAP-DIR), integrating source contours, dose, clinical text and prior-registration features. We evaluated 16,361 CT examinations from 3,473 patients across four institutional cohorts and five public benchmarks. CAP-DIR increased internal-test organ-at-risk-region normalized cross-correlation from 84.76% to 96.92% relative to the strongest learning-based comparator, with low deformation folding, and remained superior across seven external cohorts. In a blinded review of 20 longitudinal cases, clinically acceptable contours increased from 45.0% to 82.5% for clinical target volumes and from 22.5% to 62.5% for organs at risk, while estimated editing time decreased. In a controlled benchmark, CAP-DIR reduced global dose-mapping error by 34.5% relative to the strongest comparator. These findings show that clinically informed, longitudinal registration can improve contour usability and dose mapping for adaptive radiotherapy.
Tags
Links
- Source: https://arxiv.org/abs/2610.18926v1
- Canonical: https://arxiv.org/abs/2610.18926v1
Trouble viewing inline? Open PDF directly →
Full Text
86,339 characters extracted from source content.
Expand or collapse full text
Personalising Deformable Image Registration with Clinical Context and Deformation History for Adaptive Radiotherapy Caiwen Jiang 1 , Songlin Zhao 1 , Yuzhen Ding 1 , Jia Mi 3 , Samir H. Patel 1 , Alexandra Bennett 1 , Keldon K. Lin 1 , Terence T. Sio 1 , Jonathan B. Ashman 1 , Lisa A. McGee 1 , Jean-Claude M. Rwigema 1 , William G. Rule 1 , Sameer R. Keole 1 , Sujay A. Vora 1 , William W. Wong 1 , Nathan Y. Yu 1 , Michele Y. Halyard 1 , Steven E. Schild 1 , Safia K. Ahmed 1 , Daniel J. Ma 2 , Robert L. Foote 2 , Michelle A. Neben Wittich 2 , Thomas J. Whitaker 4 , Haibo Lin 5 , Dinggang Shen 3 , and Wei Liu 1,* 1 Department of Radiation Oncology, Mayo Clinic, Phoenix, Arizona, USA 2 Department of Radiation Oncology, Mayo Clinic, Rochester, Minnesota, USA 3 School of Biomedical Engineering, ShanghaiTech University, Shanghai, China 4 Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, Texas, USA 5 New York Proton Center, New York, New York, USA * Correspondence to Wei Liu (liu.wei@mayo.edu). ABSTRACT Adaptive radiotherapy must account for anatomical changes during treatment, but doing so depends on reliably mapping targets, organs at risk and dose across serial images. Existing registration methods treat each image pair independently, without using radiotherapy-specific context or the patient’s preceding anatomical trajectory. We developed Clinical-context-Aware and Patient-adaptive Deformable Image Registration (CAP-DIR), integrating source contours, dose, clinical text and prior-registration features. We evaluated 16,361 CT examinations from 3,473 patients across four institutional cohorts and five public benchmarks. CAP-DIR increased internal-test organ-at-risk-region normalized cross-correlation from 84.76% to 96.92% relative to the strongest learning-based comparator, with low deformation folding, and remained superior across seven external cohorts. In a blinded review of 20 longitudinal cases, clinically acceptable contours increased from 45.0% to 82.5% for clinical target volumes and from 22.5% to 62.5% for organs at risk, while estimated editing time decreased. In a controlled benchmark, CAP-DIR reduced global dose-mapping error by 34.5% relative to the strongest comparator. These findings show that clinically informed, longitudinal registration can improve contour usability and dose mapping for adaptive radiotherapy. Adaptive radiotherapy aims to maintain treatment accuracy as patient anatomy changes over a course of repeated treatment fractions. Tumour regression, organ filling, weight loss and other anatomical changes can alter the relationship between the planned and delivered dose 1–3 . This treatment strategy responds to these changes by acquiring treatment-state images and modifying the plan when anatomical or dosimetric deviations become clinically important 4, 5 . Deformable image registration (DIR) provides the spatial correspondence needed to propagate contours and map dose between the planning image and subsequent images. Its accuracy therefore directly affects the information available for adaptive assessment. Reliable and efficient longitudinal mapping is therefore not merely a technical component of adaptive radiotherapy but a prerequisite for translating repeated treatment imaging into clinically actionable adaptation. These demands differ from those of registration in a single diagnostic imaging study. A radiotherapy course is organized around a treatment plan, defined targets, organs at risk (OARs) and spatially heterogeneous dose objectives. Small misregistrations can be inconsequential in homogeneous soft tissue yet important near a target boundary, serial organ or steep dose gradient. The consequences also extend beyond a numerical registration score. Inaccurate propagation can increase contour-editing workload, distort structure-level dose estimates and delay decisions about whether replanning is needed. A clinically useful DIR system should therefore preserve broad anatomical correspondence while assigning particular importance to the structures and regions that determine treatment. Most DIR algorithms nevertheless formulate each registration as a generic and independent image-matching problem. Classical optimization methods can be slow in time-sensitive workflows, whereas learning-based methods such as VoxelMorph 6 , TransMorph 7 , CorrMLP 8 , MambaMorph 9 , DiffuseMorph 10 and UniGradICON 11 primarily learn population-level image correspondence. Global image similarity may underrepresent smaller or low-contrast organs at risk (OARs), although errors in these regions can alter contour review and dose assessment. Pairwise methods also discard the sequence of anatomical states already observed for the same patient. This independence assumption is especially restrictive in fractionated radiotherapy. The current image is not an unrelated sample. It follows earlier images acquired under the same treatment course, and those images may reveal patient-specific patterns such as reproducible organ filling, progressive weight loss or recurring respiratory displacement. Population-level training captures common deformation patterns, but it does not explicitly use the trajectory 1 arXiv:2610.18926v1 [eess.IV] 13 Oct 2026 already observed for an individual patient. Conversely, patient-specific optimization at inference may be too slow or unstable for routine online use. Representing preceding registrations as a fixed-length memory offers a way to use longitudinal evidence without retraining the model for every patient. The radiotherapy record provides two complementary sources of information that could address these limitations. The first is multimodal clinical context. Structure contours identify targets and OARs for which propagation errors require clinical review, dose distributions identify regions where small spatial errors can alter dose estimates, and clinical descriptions provide semantic information about the treatment site, diagnosis and plan. Together, these spatial and semantic signals define where registration accuracy is most consequential. The second is longitudinal deformation history. Earlier treatment images and registration outputs document the patient’s evolving anatomical trajectory, including recurring motion and progressive treatment-related change. Recent advances in multimodal learning 12–14 and sequential modelling 15 make it possible to use these complementary resources in a unified registration system. We therefore developed Clinical-context-Aware and Patient-adaptive Deformable Image Registration (CAP-DIR), which integrates source-side clinical context with a temporally ordered memory of previous registrations. Clinical context directs registration towards treatment-relevant anatomy, while longitudinal memory adapts the inference to the anatomical trajectory already observed for the individual patient. To evaluate CAP-DIR, we used 16,361 CT examinations from 3,473 patients across four institutional adaptive-radiotherapy cohorts and five public longitudinal imaging benchmarks, spanning seven anatomical sites and more than 20 disease categories (Fig. 1). We first assessed image similarity and deformation regularity on a patient-level internal test set and seven held-out external cohorts. We then examined anatomical and clinical relevance using longitudinal CT pairs with manual contours, a blinded clinician review of propagated contours and a controlled dose-mapping benchmark. Separate analyses quantified the contributions of clinical context and preceding patient history. All conditioning inputs were restricted to source-side or preceding information, whereas target-side contours were used only as evaluation references. This design tested whether CAP-DIR improves not only image similarity but also the anatomical correspondence and downstream tasks that support adaptive radiotherapy. I Results I.1 Study design and cohorts CAP-DIR estimates a source-to-target deformation field from a planning or reference CT and a subsequent treatment-state CT. Stage 1 conditions the registration backbone on source-side contours, dose and structured clinical text. Stage 2 additionally summarizes image features, deformation statistics and quality measures from earlier registrations for the same patient (Fig. 1a,c). Target-side contours were withheld from the model and used only for evaluation. The source-side information reflects what would be available before registering the current image. Structure contours define treatment-relevant anatomy, the dose distribution provides a spatial representation of treatment emphasis, and the clinical text describes variables such as disease site, diagnosis and treatment context. In the longitudinal stage, the memory contains only registrations that precede the current target image. The model therefore does not use future images or target-side contours. Training and model selection were performed separately for the two stages using the internal validation partition, and the selected models were then fixed for testing. The dataset comprised 3,473 patients, 5,005 plans, 16,361 CT examinations and 11,986 source–target image pairs (Fig. 1b and Table 1). Model development used patient-level partitions of the Mayo Clinic Arizona and Minnesota cohorts. These included 2,052 patients and 7,429 pairs for training, 293 patients and 1,062 pairs for validation, and 586 patients and 2,122 pairs for internal testing. MD Anderson Cancer Center, the New York Proton Center and five public benchmarks were held out for external evaluation, comprising 542 patients and 1,373 pairs. Pair construction and the derivation of clinical-context inputs for public datasets are described in the Methods. The cohorts covered seven anatomical sites and more than 20 disease categories, with different scanners, imaging protocols and longitudinal structures. In institutional cohorts with fraction-level verification imaging, a planning CT was paired with eligible subsequent verification CTs from the same plan. When a course contained multiple planning CTs but no fraction-level verification CT, the first planning CT was paired with eligible later planning CTs. For public longitudinal benchmarks, the baseline or reference CT was paired with eligible later examinations from the same patient or imaging sequence. These fixed pairs, rather than randomly redefined pairs, were used for evaluation. I.2 Internal and external registration performance On the internal test set, CAP-DIR with clinical context and patient history achieved the highest image-similarity performance, with body-region normalized cross-correlation (NCC) of 98.61%, OAR-region NCC of 96.92%, body-region structural similarity of 94.42% and OAR-region structural similarity of 93.36% (Fig. 2a). CorrMLP was the strongest general-purpose learning-based comparator, reaching body-region NCC and structural similarity of 96.08% and 89.94%, and OAR-region values of 84.76% and 88.93%, respectively. CAP-DIR therefore improved the two body-region measures by 2.53 and 4.48 percentage points and the two OAR-region measures by 12.16 and 4.43 percentage points. The clinical-context model without patient history already reached OAR-region NCC of 94.88% and structural similarity of 90.74%, with longitudinal conditioning 2/27 providing the remaining improvement. CAP-DIR also retained a low OAR folding ratio of 0.08%, compared with 0.91% for CorrMLP, and an OAR-region Jacobian deviation of 0.0038, indicating that the accuracy gain was not obtained through widespread non-physical deformation. Detailed internal results are provided in Supplementary Tables 1 and 2. The internal improvement was concentrated in the clinically specific OAR region. The body-to-OAR NCC gap ranged from 12.49 to 21.30 percentage points for TransMorph, UniGradICON, MambaMorph, VoxelMorph and DiffuseMorph and was 11.32 percentage points for CorrMLP. Clinical context reduced this gap to 3.04 percentage points, and adding patient history reduced it further to 1.69 percentage points (Fig. 2c). This regional advantage was observed across all seven anatomical sites (Fig. 2b). Relative to CorrMLP, OAR-region NCC increased by 6.0 percentage points in lung, 19.6 in pelvis, 17.8 in abdomen, 13.8 in brain and 11.1 in head-and-neck cases, whereas body-region gains ranged from 1.0 to 4.3 percentage points. The overall improvement was therefore neither limited to broad body alignment nor driven by a single anatomical cohort. Without retraining on the held-out data, CAP-DIR achieved the highest values for all four reported image-similarity measures across the seven external cohorts (Fig. 2d,e). Absolute OAR-region NCC was 95.8% in NLST, 89.5% in Longitudinal CT, 97.8% in RIDER, 82.5% in Pancreas-CT, 99.5% in 4D-Lung, 92.8% in the New York Proton Center cohort and 92.2% in the MD Anderson cohort. Gains over the strongest comparator within each cohort ranged from 0.8 to 14.3 percentage points for OAR-region NCC and from 1.2 to 11.8 percentage points for structural similarity, including OAR-region NCC gains of 6.2 and 6.3 percentage points in the New York Proton Center and MD Anderson cohorts, respectively. The greatest improvement occurred in Pancreas-CT, whereas the smaller gain in 4D-Lung reflected near-ceiling baseline performance. Although absolute performance varied with anatomy and image source, the direction of the comparison was consistent across public and institutional datasets. Cohort-level numerical results are provided in Supplementary Table 3. I.3 Contour propagation and blinded review Because improved image similarity alone does not establish clinical usefulness, we next tested whether the predicted corre- spondence produced more accurate and clinically usable contour propagation. This analysis used longitudinal planning CTs with clinician-delineated contours on both the source- and target-side images, including 276 first-to-subsequent pairs from the internal cohort and 44 additional pairs from the two private external cohorts. CAP-DIR increased OAR Dice similarity coefficient (DSC) to 92.07%, compared with 82.28% for CorrMLP, and increased clinical target volume (CTV) DSC to 87.14%, compared with 73.32% (Fig. 3a–c). Across structures and sites, this represented an overall DSC improvement of 57% relative to unregistered contour transfer, 30% relative to Demons 16 , a conventional intensity-based algorithm commonly used in clinical registration systems, and 14% relative to CorrMLP. The residual symmetric-difference volume between propagated contours and target-side reference contours decreased from 9.66 to 2.06 ml on the common 1-m grid (Fig. 3d,e), corresponding to reductions of 79%, 70% and 55% relative to direct transfer, Demons and CorrMLP, respectively. Site-specific reductions relative to direct transfer ranged from 74% to 81%. These findings provided anatomical evidence independent of the intensity metrics used for model development. Additional cohort and residual-volume details are provided in the Supplementary Note “Contour propagation and residual contour volume”. We next examined whether the quantitative improvements translated into clinically usable propagated contours. Two radiation oncology reviewers performed a method-blinded assessment of 20 longitudinal cases, with four cases each from the abdomen, brain, breast or chest wall, head and neck, and pelvis. Each case contained the clinician-approved reference contours and two unedited contour sets propagated by CAP-DIR and CorrMLP. The reference set was identified, whereas the two propagated sets were shown on the same target-side image using randomized anonymous labels. The review was performed in the clinical ARIA environment at Mayo Clinic Arizona using the image display and contour-review interface used in routine radiotherapy practice. Reviewers rated CTVs and OARs separately on a four-point scale. A score of 4 indicated clinical acceptability without modification, 3 indicated the need for minor edits, 2 indicated the need for major edits, and 1 indicated that the contours were unusable. Scores of 3 or 4 were classified as clinically acceptable. Reviewers also estimated the time required to make each result clinically usable. The proportion rated clinically acceptable increased from 45.0% with CorrMLP to 82.5% with CAP-DIR for CTVs and from 22.5% to 62.5% for OARs (Fig. 3f). Median clinician-estimated editing time decreased from 11.3 to 8.8 min for CTVs and from 26.3 to 15.0 min for OARs (Fig. 3g and Supplementary Table 4). Patient-level bootstrap confidence intervals and inter-reviewer agreement are reported in Supplementary Table 4. The largest absolute increase in clinical acceptability occurred for OARs, which often comprise several propagated structures and therefore create a greater editing burden. The agreement between the objective contour analysis and the reviewers’ assessments indicates that the anatomical improvement extended to clinical usability. I.4 Controlled dose-mapping benchmark We next assessed whether the registration improvement reduced error in accumulated-dose mapping. Because voxel-wise ground-truth accumulated dose is unavailable in clinical data, we used a controlled synthetic benchmark that generated a reference dose in the planning state and treatment-state doses under prescribed treatment-course deformations (Fig. 4a). The known deformations allowed the recovered accumulated dose to be compared voxel by voxel with ground truth. Direct transfer 3/27 produced a mean absolute error of 1.83% of the dose scale, compared with 1.43% for Demons, 1.10% for CorrMLP and 0.72% for CAP-DIR (Fig. 4b). CAP-DIR therefore reduced global dose error by 34.5% relative to CorrMLP. At the structure level, CAP-DIR reduced the average error across OAR and CTV mean- and maximum-dose endpoints from 0.97% with CorrMLP to 0.64% (Fig. 4c). CTV mean-dose error decreased from 1.10% to 0.76%, CTV maximum-dose error from 0.97% to 0.63%, OAR mean-dose error from 0.87% to 0.58%, and OAR maximum-dose error from 0.92% to 0.60%. Reductions were also observed for regional, high-gradient and dose–volume-histogram errors across anatomical sites (Fig. 4d,e). These findings isolate the effect of registration on dose mapping under known spatial transformations. The reductions in structure-level and high-gradient errors link improved spatial correspondence to the reliability of dose quantities used in adaptive assessment. They do not constitute prospective validation of clinical delivered-dose accumulation or capture all sources of uncertainty in that process. Additional site-specific results are provided in the Supplementary Note “Controlled dose-mapping results” and Supplementary Table 5. I.5 Contribution of clinical context We compared image-only registration with models conditioned on structured clinical text, spatial context from contours and dose, or both (Fig. 5a). All Stage 1 ablations used the same architecture, data partition, training schedule and OAR-weighted objective, so they differed only in the context supplied to the model. Image-only registration reached a body-region NCC of 95.8% but an OAR-region NCC of 84.4%. Structured clinical text increased OAR-region NCC to 90.1%, spatial context increased it to 92.4%, and their combination increased it to 94.9% (Fig. 5b). The combined model also increased body-region NCC to 97.9%, producing the highest performance in both regions. The distribution of the gains showed that clinical context contributed more in OARs than in the body region. Text conditioning increased body- and OAR-region NCC by 1.6 and 5.7 percentage points, respectively, whereas contours and dose increased them by 1.2 and 8.0 percentage points. Combining both sources increased body-region NCC by 2.1 percentage points and OAR-region NCC by 10.5 percentage points (Fig. 5c,d), narrowing the body-to-OAR gap from 11.4 to 3.0 percentage points. This regional benefit was observed across all anatomical sites. The combined context improved OAR-region NCC by 5.4 percentage points in lung, 6.3 in breast, 14.9 in abdomen and 16.5 in pelvis, with the largest gains in abdominal and pelvic cases (Fig. 5e). Thus, clinical context did not simply increase global image similarity but redirected registration gains towards treatment-relevant anatomy. I.6 Contribution of patient history The longitudinal memory combined features from earlier images with summaries of the predicted deformation and its registration quality. Stage 2 used only observations preceding the current target-side image and did not perform patient-specific fine-tuning at inference. We compared this history-conditioned model with the clinical-context-only model across the available history-length strata (Fig. 6). With clinical context alone, OAR-region NCC ranged from 95.5% among cases with one previous observation to 92.5% among those with 17. The history-conditioned model ranged from 97.1% to 96.0% across the same strata and improved OAR-region NCC at every evaluated history length. The history-related OAR gain ranged from 1.6 percentage points among cases with one previous observation to 3.5 percentage points among cases with 17, whereas the body-region gain ranged from 0.6 to 1.0 percentage points. Gains were 0.9–2.0 percentage points in brain and breast cases and 2.4–2.7 percentage points at shorter histories in head-and-neck and pelvic cases. Abdominal cases showed the largest site-specific benefit, ranging from 2.2 to 5.2 percentage points across the reported strata. Two patterns emerged from these results. The larger gains in OAR than in body regions indicate that longitudinal memory primarily refined structure-focused correspondence rather than uniformly increasing global alignment. Across the observed strata, the history-conditioned model maintained OAR-region NCC between 96.0% and 97.1%, whereas the clinical-context-only model ranged from 92.5% to 95.5%. The larger benefits in abdominal, head-and-neck and pelvic cases are also consistent with greater value from patient-specific history in settings characterized by recurring organ motion or progressive anatomical change. Together, these findings show that preceding registration information complemented population-level clinical context and supported more stable alignment of treatment-relevant anatomy across longitudinal imaging. The observed history-length ranges are summarized in the Supplementary Note “History-length stratification”. I Discussion In a large multi-cohort evaluation, CAP-DIR frames deformable registration in adaptive radiotherapy as a clinically conditioned and longitudinal inference problem. Across internal and external cohorts, its largest improvement occurred in OAR regions rather than in broad body regions. This regional advantage was maintained across seven external cohorts and extended beyond image-similarity measures. Evaluation against clinician-approved contours showed more accurate anatomical correspondence, the blinded review found a higher proportion of clinically acceptable propagated contours with less editing, and the controlled dose-mapping benchmark showed lower errors in accumulated-dose maps and structure-level dose summaries. The concordance 4/27 between regional registration metrics, contour propagation, blinded clinical review and controlled dose mapping indicates that the improvement extended beyond image similarity to tasks directly involved in adaptive assessment. The concentration of these gains in OAR regions is particularly relevant because whole-image similarity can be dominated by large, relatively stable regions. A model may therefore achieve a high global score while remaining inaccurate at smaller structures or interfaces where propagation errors require clinical review. CAP-DIR reduced the body-to-OAR performance gap by using radiotherapy planning information to identify where registration accuracy was most consequential. The agreement between regional image metrics, manual-contour evaluation and blinded review indicates that this conditioning changed the anatomical distribution of registration accuracy rather than merely increasing global similarity. In the clinical review, the improvement was observed for both CTVs and OARs and was particularly marked for OAR acceptability and editing requirements. These results support the use of CAP-DIR to provide a more accurate starting point for clinician review rather than to replace that review. The ablation analyses provide insight into the source of these clinically focused gains. Source contours and dose maps supplied spatial information about treatment-relevant anatomy, while structured clinical text supplied semantic information about the treatment site, diagnosis and plan. Combining these forms of clinical context produced larger OAR gains than either source alone. The temporally ordered memory then added patient-specific information from preceding images, predicted deformations and registration-quality features, improving performance beyond the clinical-context-only model across the evaluated history strata and anatomical sites. These two components also have complementary practical roles. Planning contours, dose and structured metadata can be reused from treatment preparation, whereas the memory is updated as treatment images become available. CAP-DIR can therefore adapt its inference to the observed anatomical course while retaining a fixed population-trained model and requiring neither target-side contours nor patient-specific optimization at inference. Beyond identifying the contributions of clinical context and patient history, the external evaluation and dose benchmark tested whether the resulting gains generalized across imaging settings and extended to a downstream dosimetric task. Independent radiotherapy cohorts and public longitudinal benchmarks varied in anatomy, acquisition and follow-up, producing variation in absolute performance, yet CAP-DIR maintained its relative advantage. This consistency supports transportability across imaging settings, although performance remains most appropriately interpreted within each cohort. The dose benchmark addressed a distinct consequence of spatial correspondence. Because a true patient deformation is unavailable for voxelwise validation, prescribed deformations provided a controlled reference for measuring dose-mapping error. Reductions in voxel- level, high-gradient and structure-level errors showed that the gains in registration propagated to dosimetric quantities. Taken together with the contour review, these experiments position the method as a form of clinical assistance in which propagated contours and mapped dose support assessment and adaptation while remaining subject to clinician review and quality control. These findings should nevertheless be interpreted in light of several limitations. Although CAP-DIR showed consistent performance across seven held-out external cohorts, model development was based on data from two centres within the Mayo Clinic system, and broader prospective evaluation across additional institutions and practice settings is warranted. The scope of generalisability is further bounded by the present focus on longitudinal CT-to-CT registration, leaving performance with other imaging modalities and cross-modality registration to be established. Within this evaluated setting, clinician-approved contours provided a clinically relevant anatomical reference, but they remain subject to observer variability rather than representing an absolute anatomical ground truth. Beyond these constraints on generalisability and the reference standard, the longitudinal design introduces a model-specific dependency because its memory incorporates deformation and quality features from preceding Stage 1 registrations. Errors in these earlier estimates could therefore propagate into subsequent registrations, and the robustness of the memory mechanism to such error propagation remains to be evaluated. I Conclusion CAP-DIR demonstrates that deformable image registration for adaptive radiotherapy can benefit from integrating radiotherapy- specific clinical context with longitudinal patient history. Across internal testing and seven held-out external cohorts, this formulation preferentially improved registration in treatment-relevant OAR regions while maintaining plausible deformations. The gains extended to contour propagation, blinded clinical assessment and controlled dose mapping, linking improved anatomical correspondence with tasks central to adaptive radiotherapy. These findings support clinically informed and patient- adaptive registration as a means of providing more reliable starting contours and longitudinal dose estimates for adaptive assessment across the treatment course. IV Methods IV.1 Ethics This multi-site retrospective study was approved by the Mayo Clinic Institutional Review Board (IRB #: 24-011106), which served as the single IRB of Record for the participating sites, and was conducted in accordance with the Declaration of Helsinki. 5/27 Given the retrospective design, the Mayo Clinic Institutional Review Board waived the requirement for study-specific written informed consent. Data transfer and analysis were performed under the corresponding institutional data-use arrangements. Publicly available de-identified datasets were used in accordance with their original access and data-use requirements. IV.2 Study datasets and preprocessing Longitudinal radiotherapy cohorts. Institutional longitudinal radiotherapy data were collected from Mayo Clinic Arizona, Mayo Clinic Minnesota, MD Anderson Cancer Center and the New York Proton Center, and public longitudinal CT benchmarks were included to broaden the imaging protocols and anatomical settings represented in evaluation. Across cohorts, available information included CT images acquired across the treatment course, structure contours, dose distributions and patient-level clinical text. For public datasets that did not contain complete clinical information across the treatment course, the clinical- context inputs required by the model were completed using a standardised preprocessing pipeline. Structured clinical text was assembled from available de-identified dataset annotations, DICOM header fields and imaging metadata. Source structure masks were obtained from available baseline or reference annotations and supplemented, when necessary, using the fixed pretrained TotalSegmentator model 17 . When clinical dose distributions were unavailable, source-space proxy dose maps were generated from the available source target contours using a standardised geometric distance-to-target formulation 18 . All derived clinical-context inputs were constructed from source-side information, while target annotations were reserved exclusively for evaluation. All records were organised by patient, treatment course and treatment plan to preserve the longitudinal relationship among plans, fractions and adaptive imaging time points. Detailed cohort composition, patient-level partitioning and image-pair counts are provided in Table 1. Preprocessing. Preprocessing was performed separately for imaging data and clinical text. For CT examinations from the same patient, planning or reference CTs and verification or treatment-state CTs were first reoriented to a common anatomical orientation and harmonised in physical space to account for differences in image origin, direction, matrix size, voxel spacing and field of view. CT volumes were then resampled to an isotropic resolution of1× 1× 1m 3 and cropped or padded to a fixed matrix size of512× 512× 256. CT intensities were clipped to a fixed HU range of[−1000, 1000]to retain air, soft-tissue and high-density anatomical information relevant to radiotherapy registration, and were linearly normalised to[−1, 1]. Structure masks were transformed using the same spatial operations with nearest-neighbour interpolation. Dose maps derived from the treatment-planning dose grid were first mapped to the physical coordinate system of the corresponding CT and then subjected to the same resampling, cropping and padding operations as the CT volume using linear interpolation. Dose values were kept in Gy during preprocessing, and voxels outside the original dose grid were assigned zero dose. Clinical text was extracted from CT planning records, treatment-plan information and clinical notes, and then organised into a structured template before model training. The template summarised the anatomical site, diagnosis, treatment plan, prescribed dose and fractionation, image-acquisition information and relevant structure context, including available OAR and CTV information. IV.3 Image-pair construction and dataset partitioning Throughout the primary registration analyses, a registration case denotes one source–target CT pair together with the auxiliary inputs available to the evaluated model. For CAP-DIR Stage 1, these inputs comprise the image pair and its corresponding clinical context; Stage 2 additionally includes the patient’s available prior treatment history. The source–target pair uniquely defines the case, so adding clinical context or history does not change the case count. Image-pair construction depended on the imaging structure available in each cohort. For cohorts with fraction-level verification imaging, each planning CT was paired with every eligible subsequent verification CT from the same treatment plan. For cohorts in which each plan contained only one planning CT and no fraction-level verification CT, plans were ordered chronologically within each patient and treatment course, and the planning CT from the first plan was paired with each eligible planning CT from subsequent plans. For public longitudinal imaging benchmarks without explicit plan–fraction structure, the baseline or reference CT was paired with eligible later CT examinations from the same patient or imaging sequence. The resulting dataset contained 3,473 patients, 5,005 treatment plans, 16,361 CT examinations and 11,986 image pairs. Dataset-level counts are reported in Table 1. Mayo AZ and Mayo MN patients were randomly split at the patient level into training, validation and internal test sets using a 70%, 10% and 20% split, respectively. All scans from the same patient were assigned to only one split to prevent patient-level data leakage. This produced 2,052 training patients with 7,429 image pairs, 293 validation patients with 1,062 image pairs and 586 internal test patients with 2,122 image pairs. MD Anderson, NYPC and all five public datasets were held out entirely for external validation, comprising 542 patients and 1,373 image pairs. Within these fixed patient-level splits, Stage 1 treated the image pairs as independent registration cases. For Stage 2, samples were organised as a current image pair together with a temporally ordered sequence of previous image pairs from the same patient. Only prior fractions or plans were included in the history for a given current target image, preventing leakage from the current or future time points. During training, current cases were randomly drawn from the corresponding fixed split. 6/27 IV.4 CAP-DIR framework The core CAP-DIR network estimates a dense source-to-target deformation field from a source CTI s and target CTI t (Fig. 7a). The two CT volumes are processed by weight-shared 3D CNN encoders to extract hierarchical image features. Each encoder contains five resolution levels, with two3× 3× 3convolutional layers, instance normalisation and LeakyReLU activation at each level, followed by downsampling to construct a feature pyramid. Corresponding source and target features are fused at each resolution, forming multi-scale representationsF l 5 l=1 for deformation refinement. The fused feature pyramid is passed to a progressive transformer decoder. From the coarsest to the finest resolution, each decoder level models long-range contextual dependencies with a Swin Transformer block and predicts a residual displacement field∆φ l . The deformation field from the previous coarser level is upsampled and additively composed with the residual update: φ l = Up(φ l+1 )+∆φ l .(1) This progressive refinement provides a stable coarse initialisation for large deformations while preserving finer anatomical detail at higher resolutions. The final deformation fieldφ s→t is used to warp the source CT by trilinear interpolation, enabling image-similarity loss computation and registration assessment. The multimodal clinically informed stage (Stage 1) and the patient-adaptive stage (Stage 2) use the same image-registration backbone. CAP-DIR was therefore formulated as a conditional registration framework. Stage 1 conditions the shared registration backbone on the current multimodal clinical context, whereas Stage 2 further conditions the same backbone on longitudinal patient-history information through a memory encoder. Clinical-context conditioning. All clinical-context inputs, including structure contours, dose maps and planning text, were restricted to source-side information available before registration. Manual contours on target CTs, when available, were withheld from the model and used exclusively for evaluation. As shown in Fig. 7b,c, the clinical-context module used two complementary sources of information from the radiotherapy workflow. Semantic context was provided by the structured text prompt, which summarised the anatomical site, diagnosis, treatment plan, prescribed dose and fractionation, image-acquisition information and relevant structure context. Spatial context was provided by structure contours and dose maps, which indicate anatomy and treatment-sensitive regions where registration errors are more clinically consequential. These semantic and spatial inputs were converted into decoder conditioning signals as follows. As illustrated in Fig. 7b, semantic context was first standardised into a structured text prompt before being passed to the registration network. This step converted heterogeneous clinical records into a consistent representation of the treatment site, diagnosis, plan information, dose prescription and relevant structure context, reducing variation caused by free-text formatting rather than clinical content. The structured prompt was then encoded using a frozen BioMedBERT-large encoder 19 , a BERT-based language model pretrained on large-scale biomedical text, to obtain a domain-specific semantic representation of the current registration sample. The resulting 768-dimensional embedding e was passed through a two-layer prompt adapter with GELU activation. The adapter converted the semantic embedding into feature-wise linear modulation parameters, allowing clinical information to condition the decoder without replacing the image-driven registration backbone. For decoder levels l∈1, 2, 3, 4, 5, the modulation parameters were defined as: (γ l , β l ) = Linear l PromptAdapter(e) .(2) Spatial context was derived from available source-side structure contours and dose maps for the current treatment plan. The goal was to provide the decoder with spatial priors that indicate not only where clinically important anatomy is located, but also where registration errors are more likely to affect adaptive assessment. As illustrated in Fig. 7c, CTV contours were converted into signed-distance maps and transformed into Gaussian proximity and boundary-band features, preserving both the interior of the CTV and its clinically sensitive boundary. OAR contours were projected into binary structure maps to identify organs in which contour-transfer or dose-mapping errors would require clinical review. These contour-derived features were fused to form a continuous anatomical proximity mapA∈ [0, 1] H×W×D . Dose maps were normalised and processed to extract isodose-band and local dose-gradient features, because small spatial errors near steep dose gradients can lead to disproportionate dosimetric errors. These dose-derived features were combined with the OAR map to emphasise dose-sensitive anatomy, yielding a dose-sensitive map ˆ D∈ [0, 1] H×W×D . The anatomical proximity map and dose-sensitive map were then combined to form a risk map: R = 1−(1−A)(1− ˆ D).(3) The spatial priorP = concat(A,R) was downsampled to each decoder scale l∈1, 2, 3, 4, 5 and used as a spatial gate: ̃ F l = F l ⊙ σ W l [F l , Pool l (P)]+ b l ,(4) 7/27 whereσdenotes the sigmoid function and⊙denotes element-wise multiplication. Semantic modulation was then applied to the spatially gated decoder feature map: ˆ F l = γ l ⊙ ̃ F l + β l .(5) Patient-history conditioning. Patient-history conditioning was used only in Stage 2, while the current CT pair and current multimodal clinical-context inputs from Stage 1 were retained. For a current registration, the available patient history included previous treatment-state images, previous Stage 1 deformation outputs and registration-quality information from the same patient. Together, these sources describe the anatomical states already observed, the patient’s prior motion and shape-change pattern, and the reliability of each historical registration when updating the patient-specific memory. To incorporate these sources into the current registration, Stage 2 converted them into the memory representation shown in Fig. 7d. Directly processing the original historical CT volumes and full dense deformation fields would substantially increase memory and computation and would make variable-length patient histories difficult to handle. Therefore, each previous time point was represented by a compact memory feature. After Stage 1 training, the trained Stage 1 model was applied to each previous source-target CT pair from the same patient, and the resulting image features, deformation-field statistics and quality measurements were stored as the history representation. For a patient withhprevious time points, the history sequence was indexed in temporal order as1,..., h. For each historical registration, the Stage 1 encoder produced a 128-dimensional image feature vector v t by global average pooling of bottleneck features. Deformation memory was represented by four statistics, including mean displacement magnitude, 95th-percentile displacement, negative Jacobian fraction and Jacobian standard deviation, forming d t ∈R 4 . These statistics summarised the magnitude and regularity of the previous deformation without storing the full dense vector field. Quality memory acted as a confidence summary and included NCC body , NCC OAR , SSIM body , SSIM OAR and mean inverse-consistency error, forming q t ∈R 5 . Historical OAR-specific quality measurements were computed within source-derived OAR masks warped into target space using the corresponding Stage 1 deformation fields, without access to manual target contours. The per-time-point memory feature was f t = [v t ;d t ;q t ]∈R 137 .(6) A trend feature was computed from successive deformation statistics: τ t = d t −d t−1 ,(7) withτ 1 set to zero. The sequence(f 1 ,..., f h ) was passed to a single-layer GRU with hidden dimension 256. The GRU was used because it can summarise temporally ordered histories with variable length while preserving the order of previously observed anatomical changes, yielding the patient-history embedding m = GRU(f 1 ,...,f h ) h .(8) The history embedding was concatenated with the current source CT feature p∈R 128 and current trend feature τ h : z = [p;m; τ h ]∈R 388 .(9) A two-layer memory adapter mapped z to patient-adaptive FiLM parameters(γ ∗ l , β ∗ l ) for the same decoder levelsl∈1, 2, 3, 4, 5 used by the clinical-context module. These parameters provided patient-specific conditioning in addition to the current clinical- context conditioning. Two-stage training and objective. CAP-DIR was trained sequentially. Stage 1 was trained first using independent current source-target CT pairs and the corresponding source-derived clinical-context inputs, including structured text prompts, structure contours and dose maps. In this stage, the shared registration backbone and clinical-context conditioning modules were optimised to estimate φ s→t for the current image pair. After Stage 1 convergence, historical representations were generated for the training patients as described above. Stage 2 was then trained with current-pair-plus-history sequences. For each current target image, the input included the current CT pair, current clinical context and only previous time points from the same patient. During Stage 2 training, the Stage 1 registration backbone was initialised from the trained Stage 1 model. The image encoder was kept frozen, whereas the decoder, GRU memory encoder, memory adapter and patient-adaptive modulation parameters were optimised. During inference, patient history was used only as input to the trained memory encoder, and no patient-specific fine-tuning or test-time parameter update was performed. Both stages used the same image-similarity and regularisation objective: L =L NCC + λ OAR L OAR NCC + λ reg L reg .(10) 8/27 L NCC andL OAR NCC denote local NCC losses computed in the foreground body and OAR regions, respectively. ForL OAR NCC , the source-derived OAR mask was warped into target space using the deformation field predicted for the current image pair. The resulting mask provided a coarse target-space OAR localisation for computing the OAR-weighted NCC loss, without using manual target contours. The foreground body mask was used to prevent similarity optimisation from being dominated by padded background voxels, while the OAR-weighted term increased the contribution of clinically important structures.L reg is a diffusion regularisation term on the spatial gradient of the displacement field. Unless otherwise specified,λ OAR = 1.0and λ reg = 0.2. IV.5 Implementation and comparison methods Implementation details. All models were implemented in PyTorch 2.5.1 with CUDA 11.8. CAP-DIR training was run with distributed data parallel optimisation on3×NVIDIA A100-80G GPUs. Stage 1 used a batch size of 1 per GPU, 2,000 optimisation steps per epoch. Optimisation was performed with Adam using a fixed learning rate of10 −4 andλ reg = 0.2. The clinical-text embedding dimension was 768, the prompt-modulation scale was initialised to 0.1, and the spatial-prior weight was initialised to 0.1. These two scaling factors allowed text-based modulation and contour-dose spatial priors to enter the decoder gradually at the beginning of training, rather than dominating the image-matching objective from the first optimisation steps. Stage 2 was initialised from the Stage 1 checkpoint and optimised with the same learning rate, batch size and regularisation weight. The Stage 1 and Stage 2 checkpoints used for evaluation were selected independently according to their best performance on the internal validation set. Historical information was encoded with a GRU memory aggregator, and samples with longer histories were up-weighted using a curriculum factor of1+ 0.1h, wherehis the number of available previous time points. This weighting encouraged the memory encoder to learn from patients with richer longitudinal histories while still retaining short-history samples during training. All CAP-DIR variants and learning-based comparison methods used the same preprocessed image pairs and patient-level data splits. Comparison methods. CAP-DIR was compared with representative learning-based and conventional DIR methods. Voxel- Morph 6 is a convolutional unsupervised registration network. TransMorph 7 uses a transformer architecture to model long-range correspondence. CorrMLP 8 introduces correlation-aware MLP blocks for dense displacement estimation. MambaMorph 9 uses state-space sequence modelling for medical image registration. DiffuseMorph 10 formulates registration through a diffusion-based deformation generation process. UniGradICON 11 is a general-purpose pretrained registration model based on inverse-consistency regularisation. UniGradICON was used directly without retraining, whereas the other learning-based baselines were trained using the same preprocessed image pairs and patient-level data partition as CAP-DIR. Direct transfer without registration was included to quantify the uncorrected anatomical discrepancy. Demons registration was also included because it is a conventional intensity-based DIR algorithm commonly used in clinical image-processing systems, providing a reference for routine clinical registration performance. CAP-DIR variants were used for ablation analyses. The image-only variant used the CT image pair without clinical-context conditioning. The clinical-context variant added semantic and spatial clinical information. The full model added patient-history adaptation to the clinical-context model and is denoted Ours w/ Clin.+Hist. in the figures. All Stage 1 clinical-context ablations used the same architecture, training schedule, data partition and OAR-weighted objective, differing only in the clinical-context inputs provided to the model. IV.6 Blinded clinical contour review A retrospective method-blinded clinical review was conducted using 20 longitudinal planning-CT cases, with four cases each from the abdomen, brain, breast/chest wall, head and neck, and pelvis. Each case included the clinician-approved target contour set and two unedited contour sets propagated by CAP-DIR and CorrMLP. The clinician-approved set was identified as the reference, whereas the identities of the two propagation methods were concealed using case-specific anonymous labels. All CT images and contour sets were imported into the clinical ARIA environment at Mayo Clinic Arizona. Two independent radiation oncology reviewers inspected the contours using the image display and contour-review interface used in routine radiotherapy practice, thereby approximating the clinical environment in which propagated contours would be reviewed and edited. Reviewers rated CTVs and OARs separately on a four-point ordinal scale: 4, clinically acceptable without modification; 3, requiring minor edits; 2, requiring major edits; and 1, unusable. The reviewers also estimated the time required to bring the CTV and OAR contours to clinical usability. All three contour sets were resampled to a common2.5× 2.5× 4-m grid before review, and neither automatically propagated set underwent manual correction before assessment. IV.7 Evaluation metrics Registration performance was evaluated with image-similarity, deformation-field and displacement metrics. Image similarity was measured with local normalised cross-correlation (NCC) using a9× 9× 9window and structural similarity index (SSIM). Deformation-field plausibility was assessed using Jacobian determinant statistics, including Jacobian deviation, and folding 9/27 ratio, defined as the fraction of voxels with a negative Jacobian determinant. Deformation magnitude was measured as the mean displacement-field magnitude. Contour propagation was evaluated with Dice similarity coefficient (DSC) and residual symmetric-difference volume. DSC measured the overlap between propagated contours and manual target contours. Residual volume was calculated from the symmetric difference between each propagated contour and the manual target contour and reported in millilitres; it is a geometric error measure and not a direct measurement of clinician editing effort. For the blinded clinical review, clinical acceptability was defined for analysis as a rating of 3 or 4, indicating that a contour required no more than minor edits. The clinician-based endpoints were the proportion of contours meeting this threshold and the estimated editing time for CTVs and OARs. The absolute difference in clinical acceptability and the paired reduction in estimated editing time were used to quantify the benefit of CAP-DIR relative to CorrMLP. Dose accumulation was evaluated with dose mean absolute error (dose MAE) and dose-volume histogram (DVH) endpoint error. Dose MAE measured the voxel-wise difference between estimated and reference accumulated dose. DVH endpoint error measured differences in clinically relevant dose-summary endpoints, including Dmean and Dmax. IV.8 Statistical analysis All quantitative analyses were performed on matched evaluation units, so that each method was evaluated on the same registration case, contour-propagation case or dose-accumulation case. Continuous metrics are reported as mean values with standard deviations unless otherwise specified. For metrics where higher values indicate better performance, relative improvement was calculated as the difference from the comparator divided by the comparator value. For error metrics where lower values indicate better performance, relative reduction was calculated as the comparator value minus the CAP-DIR value, divided by the comparator value. Paired comparisons between methods used two-sided Wilcoxon signed-rank tests when statistical significance was assessed. Anatomical-site and external cohort analyses were computed within each stratum using the matched cases available for that stratum. Clinical-review outcomes were analysed as paired observations matched by case and reviewer. Ordinal ratings are reported as counts and proportions, with the primary presentation based on the clinical-acceptability threshold. The two reviewer estimates of editing time were averaged within each case before reporting the median and interquartile range across cases. Absolute and relative reductions were calculated from these case-level summaries. Confidence intervals for differences in clinical acceptability were obtained by patient-level bootstrap resampling. Inter-reviewer agreement for ordinal ratings was assessed using linearly weighted Cohen’s κ . Data Availability The institutional radiotherapy cohorts used in this study are not publicly available because they contain protected clinical imaging and treatment information covered by patient privacy regulations, institutional review board approval and data-use agreements. Subject to future institutional approval, a small number of fully de-identified example cases may be released or shared through a controlled request process to support reproducibility. The public benchmark datasets used for external validation are available from their original release pages, including LongCT-Lung, NLST-New-lesion-LongCT, Pancreas-CT, 4D-Lung and RIDER Lung CT. Users should comply with the access conditions and data-use policies of each source dataset. Code Availability The code for data preprocessing, model training, testing and evaluation will be released on GitHub upon publication. The repos- itory will include the preprocessing scripts, CAP-DIR training code, baseline training and evaluation pipelines, configuration files and instructions for reproducing the main experiments. Trained model parameters derived from the institutional datasets will not be publicly released because of institutional data-governance requirements. Acknowledgements This work was supported in part by the Department of Radiation Oncology, Mayo Clinic, Rochester, Minnesota, and Phoenix, Arizona, USA, R01 CA 261932 and R01 CA 272602 to Robert W. Mutter, and P30 CA015083 to MCCCC. This research was also supported by NIH/NIBIB R01EB293388, by NIH/NCI R01CA280134, by the Eric & Wendy Schmidt Fund for AI Research & Innovation, and by the Kemper Marley Foundation to Wei Liu. Author contributions statement C.J. and W.L. conceived the study. C.J., S.Z., Y.D. and M.J. contributed to method development, software implementation, data preparation, experiments and analysis. A.B. and K.K.L. performed the blinded clinical contour review. S.H.P., T.T.S., J.B.A., 10/27 L.A.M., J.-C.M.R., W.G.R., S.R.K., S.A.V., W.W.W., N.Y.Y., M.Y.H. and S.E.S. provided clinical expertise and contributed to study design and interpretation of the results. D.S. and W.L. supervised the study, and W.L. provided project administration and funding support. C.J. prepared the original manuscript. All authors reviewed and edited the manuscript and approved the final version. Competing interests The authors declare no competing interests. References 1. Sonke, J.-J., Aznar, M. & Rasch, C. Adaptive radiotherapy for anatomical changes. Semin. Radiat. Oncol. 29, 245–257, DOI: 10.1016/j.semradonc.2019.02.006 (2019). 2.Glide-Hurst, C. K. et al. Adaptive radiation therapy (ART) strategies and technical considerations: A state of the ART review from NRG oncology. Int. J. Radiat. Oncol. Biol. Phys. 109, 1054–1075, DOI: 10.1016/j.ijrobp.2020.10.021 (2021). 3.Hall, W. A. et al. Magnetic resonance linear accelerator technology and adaptive radiation therapy: An overview for clinicians. CA: A Cancer J. for Clin. 72, 34–56, DOI: 10.3322/caac.21707 (2022). 4.Yan, D., Vicini, F., Wong, J. & Martinez, A. Adaptive radiation therapy. Phys. Medicine Biol. 42, 123–132, DOI: 10.1088/0031-9155/42/1/008 (1997). 5.Lavrova, E. et al. Adaptive radiation therapy: A review of CT-based techniques. Radiol. Imaging Cancer 5, e230011, DOI: 10.1148/rycan.230011 (2023). 6.Balakrishnan, G., Zhao, A., Sabuncu, M. R., Guttag, J. & Dalca, A. V. VoxelMorph: A learning framework for deformable medical image registration. IEEE Transactions on Med. Imaging 38, 1788–1800, DOI: 10.1109/TMI.2019.2897538 (2019). 7.Chen, J. et al. TransMorph: Transformer for unsupervised medical image registration. Med. Image Analysis 82, 102615, DOI: 10.1016/j.media.2022.102615 (2022). 8.Meng, M., Feng, D., Bi, L. & Kim, J. Correlation-aware coarse-to-fine MLPs for deformable medical image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9645–9654, DOI: 10.1109/CVPR52733.2024.00921 (2024). 9.Guo, T. et al. MambaMorph: a Mamba-based framework for medical MR-CT deformable registration. arXiv preprint arXiv:2401.13934 DOI: 10.48550/arXiv.2401.13934 (2024). 10.Kim, B., Han, I. & Ye, J. C. DiffuseMorph: Unsupervised deformable image registration using diffusion model. In Proceedings of the European Conference on Computer Vision (ECCV), vol. 13662 of Lecture Notes in Computer Science, 347–364, DOI: 10.1007/978-3-031-20086-1_20 (Springer, Cham, 2022). 11.Tian, L. et al. uniGradICON: A foundation model for medical image registration. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, vol. 15002 of Lecture Notes in Computer Science, 749–760, DOI: 10.1007/978-3-031-72069-7_70 (Springer, Cham, 2024). 12.Perez, E., Strub, F., de Vries, H., Dumoulin, V. & Courville, A. FiLM: Visual reasoning with a general conditioning layer. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32, DOI: 10.1609/aaai.v32i1.11671 (2018). 13.Singhal, K. et al.Large language models encode clinical knowledge.Nature 620, 172–180, DOI: 10.1038/ s41586-023-06291-2 (2023). 14. Moor, M. et al. Foundation models for generalist medical artificial intelligence. Nature 616, 259–265, DOI: 10.1038/ s41586-023-05881-4 (2023). 15. Cho, K. et al. Learning phrase representations using RNN encoder-decoder for statistical machine translation. In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), 1724–1734, DOI: 10.3115/v1/D14-1179 (Association for Computational Linguistics, 2014). 16.Thirion, J.-P. Image matching as a diffusion process: an analogy with Maxwell’s demons. Med. Image Analysis 2, 243–260, DOI: 10.1016/S1361-8415(98)80022-4 (1998). 17. Wasserthal, J. et al. TotalSegmentator: Robust segmentation of 104 anatomic structures in CT images. Radiol. Artif. Intell. 5, e230024, DOI: 10.1148/ryai.230024 (2023). 18.Appenzoller, L. M., Michalski, J. M., Thorstad, W. L., Mutic, S. & Moore, K. L. Predicting dose–volume histograms for organs-at-risk in IMRT planning. Med. Phys. 39, 7446–7461, DOI: 10.1118/1.4761864 (2012). 11/27 19.Gu, Y. et al. Domain-specific language model pretraining for biomedical natural language processing. ACM Transactions on Comput. for Healthc. 3, 1–23, DOI: 10.1145/3458754 (2021). 12/27 Figures a b c Figure 1. Overview of clinical motivation, experimental data and method design in this study. a, The longitudinal ART workflow from treatment planning to fraction-level verification imaging, DIR-based contour propagation, delivered-dose assessment and adaptation decisions. b, The large-scale longitudinal radiotherapy dataset used for CAP-DIR development and evaluation, including multi-institutional cohorts, public benchmarks, anatomical coverage and patient composition. c, The two-stage CAP-DIR workflow integrating clinically informed registration with patient-adaptive memory from previous Stage 1 outputs. 13/27 a bc d e Figure 2. CAP-DIR improves OAR-focused registration performance across internal and external cohorts. a, Quantitative comparison across image similarity and deformation regularity metrics on the internal test set. b, Body- and OAR-region gains over the strongest overall comparison method across anatomical sites. c, Relationship between Body-to-OAR performance drop and estimated OAR displacement magnitude. d, External validation heatmaps comparing all methods across held-out cohorts. e, OAR-region gains over the strongest comparison method across external validation cohorts. 14/27 f g Figure 3. Improved registration translates into more accurate contour propagation. a, Evaluation design using longitudinal planning CT pairs with manual contours available on both source and target images. b, Overall contour DSC for direct transfer, Demons, CorrMLP and CAP-DIR. c, OAR- and CTV-specific contour DSC. d, Residual symmetric-difference volume across anatomical sites; 10 3 voxels on the 1-m 3 evaluation grid correspond to 1 ml. f, Proportion of CTV and OAR contours rated clinically acceptable (no more than minor edits) in the blinded clinical review; each method comprised 40 reader–case evaluations. g, Median clinician-estimated editing time, with interquartile ranges, after averaging the two reviewer estimates within each case. e, Representative OAR and CTV propagation examples across anatomical sites. 15/27 Figure 4. Improved registration supports dose accumulation evaluation. a, Evaluation design for simulated dose accumulation. A prescribed treatment-course deformation generates a treatment-state CT and dose from a reference CT-dose pair. Reference accumulation adds the reference dose to the treatment-state dose mapped by the prescribed inverse deformation, whereas estimated accumulation adds the reference dose to the treatment-state dose mapped by direct transfer or DIR-predicted deformation. b, Overall accumulated-dose mean absolute error (MAE) for direct transfer without registration, Demons, CorrMLP and CAP-DIR. c, Absolute DVH endpoint error for OAR and CTV Dmean and Dmax. d, Anatomical-site consistency of CAP-DIR error reduction relative to CorrMLP across global, structure-region, gradient-boundary and DVH endpoint metrics. e, Representative head-and-neck and lung visualisations comparing reference dose, treatment-state dose, reference accumulation and accumulated dose estimated by each registration method. 16/27 Figure 5. Multimodal clinical context improves registration and redirects gains to clinically relevant regions. a, Ablation design comparing image-only registration with semantic clinical notes, spatial clinical context, and their combination. b, Body and OAR-region NCC across clinical-context variants. c, Body-to-OAR NCC gap across the same variants. d, Body and OAR-region NCC gains over the image-only model, highlighting preferential improvement in clinically important regions. e, Anatomical-site heatmap of OAR-region NCC gain relative to image-only registration. 17/27 Figure 6. Patient-history adaptation improves registration as longitudinal evidence accumulates. a, Patient history available for a current treatment fraction, including image, deformation-field, and registration-quality memories from previous fractions, which are summarised by a memory encoder. b, OAR-region NCC as a function of the number of prior fractions for CAP-DIR with clinical context alone and with clinical context plus patient history. c, NCC gain from patient history in the Body and OAR regions across increasing history lengths. d, Anatomical-site heatmap of OAR-region NCC gain from patient history across history lengths. 18/27 Figure 7. CAP-DIR conditional registration framework. a, Overall CAP-DIR architecture. Source and target CTs are encoded by weight-shared 3D CNNs, fused as multi-scale features and decoded by a coarse-to-fine transformer to estimate a dense deformation field for source CT warping. Multimodal clinical context conditions Stage 1, and patient-history information provides additional conditioning in Stage 2. b, Semantic clinical-context conditioning converts structured clinical text into BioMedBERT embeddings and FiLM parameters that modulate decoder features. c, Spatial clinical-context conditioning converts OAR/CTV contours and dose maps into anatomy and risk maps that gate decoder features. d, Patient-history conditioning encodes previous image features, deformation statistics and quality memory with a GRU to generate patient-adaptive FiLM parameters. 19/27 Tables Table 1. Dataset split overview. Patient, treatment-plan, CT-examination and image-pair counts are reported for each cohort across the training, validation, internal test and external validation splits. SplitDatasetInstitution / SourcePatientsPlansTotal CTsImage pairs Training Mayo AZMayo Clinic Arizona1,7562,4729,4917,019 Mayo MNMayo Clinic Minnesota296547547410 Subtotal2,0523,01910,0387,429 Validation Mayo AZMayo Clinic Arizona2513531,3561,003 Mayo MNMayo Clinic Minnesota42787859 Subtotal2934311,4341,062 Internal test Mayo AZMayo Clinic Arizona5017062,7112,005 Mayo MNMayo Clinic Minnesota85156156117 Subtotal5868622,8672,122 Internal total2,9314,31214,33910,613 External validation MD AndersonMD Anderson Cancer Center30535323 NYPCNew York Proton Center25464621 LongCT-LungPublic300331684353 NLSTPublic118122244122 Pancreas-CTPublic404012080 4D-LungPublic2080800720 RIDERPublic9217554 Subtotal5426932,0221,373 Grand total3,4735,00516,36111,986 20/27 Supplementary Information: Personalising Deformable Image Registration with Clinical Context and Deformation History for Adaptive Radiotherapy Caiwen Jiang 1 , Songlin Zhao 1 , Yuzhen Ding 1 , Jia Mi 3 , Samir H. Patel 1 , Alexandra Bennett 1 , Keldon K. Lin 1 , Terence T. Sio 1 , Jonathan B. Ashman 1 , Lisa A. McGee 1 , Jean-Claude M. Rwigema 1 , William G. Rule 1 , Sameer R. Keole 1 , Sujay A. Vora 1 , William W. Wong 1 , Nathan Y. Yu 1 , Michele Y. Halyard 1 , Steven E. Schild 1 , Safia K. Ahmed 1 , Daniel J. Ma 2 , Robert L. Foote 2 , Michelle A. Neben Wittich 2 , Thomas J. Whitaker 4 , Haibo Lin 5 , Dinggang Shen 3 , and Wei Liu 1,* 1 Department of Radiation Oncology, Mayo Clinic, Phoenix, Arizona, USA 2 Department of Radiation Oncology, Mayo Clinic, Rochester, Minnesota, USA 3 School of Biomedical Engineering, ShanghaiTech University, Shanghai, China 4 Department of Radiation Physics, The University of Texas MD Anderson Cancer Center, Houston, Texas, USA 5 New York Proton Center, New York, New York, USA * Correspondence to Wei Liu (liu.wei@mayo.edu). Supplementary Notes Contour propagation and residual contour volume The contour-propagation analysis used longitudinal planning CTs because routine verification CTs rarely contained complete manual reference contours. The first planning CT was used as the source and each eligible subsequent planning CT as a target. The internal cohort provided 276 source–target pairs; MD Anderson and the New York Proton Center provided 23 and 21 additional pairs, respectively. Direct transfer produced the greatest residual symmetric-difference volume, which CAP-DIR reduced from 9.66 to 2.06 ml. This corresponded to reductions of 79% relative to direct transfer and 55% relative to CorrMLP. Controlled dose-mapping results The dose-mapping benchmark used prescribed synthetic treatment-course deformations to provide a known reference mapping. The reference accumulated dose combined the reference dose with the treatment-state dose mapped by the prescribed inverse deformation. Estimated accumulation used direct transfer or the deformation predicted by each DIR method. CAP-DIR reduced whole-volume dose mean absolute error by 34.5% relative to CorrMLP and reduced the mean error across OAR and CTV mean- and maximum-dose endpoints by 33.3%. Across anatomical sites, the reduction relative to CorrMLP was 27.7–35.9% for global error, 26.6–36.0% within structure regions, 28.9–36.3% at gradient boundaries and 30.3–38.1% for dose-volume-histogram endpoints. History-length stratification Among cases stratified by the amount of available prior history, the OAR-region gain from Stage 2 increased from 1.6 percentage points with one prior observation to 3.5 percentage points with 17 prior observations; the body-region gain ranged from 0.6 to 1.0 percentage points. Abdominal cases showed the largest observed site-specific increase. Supplementary Tables 1 21/27 Supplementary Table 1. Quantitative comparison of registration methods on the internal test set (586 patients, 2,122 registration cases from Mayo AZ and Mayo MN). One registration case corresponds to one source–target CT pair; clinical context and available patient history are model inputs and do not alter the case count. All metrics are reported as mean± s.d. Image similarity: NCC and SSIM (%) computed over whole-body (Body) and OAR-masked (OAR) regions. Deformation regularity: mean Jacobian determinant (closer to 1.0 indicates better local volume preservation) and folding ratio (percentage of voxels with a non-positive Jacobian determinant; lower is better), each computed over Body and OAR regions. Deformation magnitude: mean displacement field magnitude (m) over Body and OAR regions, reported descriptively without ranking. Initial (no reg.): before-registration image similarity; deformation metrics not applicable (—). Bold = best per metric; underline = second best (displacement columns excluded from ranking). CAP-DIR (Clin.) uses clinical context, and CAP-DIR (Clin.+Hist.) additionally uses patient history. Both variants are evaluated on the full internal test set; for cases without prior fraction history, CAP-DIR (Clin.+Hist.) falls back to Stage 1 behaviour. Method Image SimilarityDeformation RegularityDeformation Magnitude BodyOARBodyOARBody / OAR NCC(%)↑SSIM(%)↑NCC(%)↑SSIM(%)↑Jac→1Fold%↓Jac→1Fold%↓Disp. (m)Disp. (m) Initial (no reg.)85.58±13.9773.60±10.3860.87±29.3672.48±15.99— VoxelMorph 1 84.91±8.4284.91±8.7875.46±25.8482.36±13.520.9886±0.01120.16±0.221.0148±0.04850.58±0.344.03±4.183.74±5.15 TransMorph 2 96.41±3.5289.76±5.6883.92±18.7488.61±8.661.0485±0.03461.38±1.721.0712±0.05941.84±1.895.42±6.325.18±6.21 CorrMLP 3 96.08±5.2189.94±6.0884.76±17.6288.93±8.411.0347±0.02150.62±1.311.0463±0.05280.91±1.485.36±4.824.62±4.76 DiffuseMorph 4 90.24±11.3578.82±9.7168.94±25.1276.81±13.960.9991±0.00860.04±0.081.0006±0.01210.11±0.192.90±0.532.30±0.83 MambaMorph 5 96.25±5.8687.86±7.4281.28±20.6686.32±10.081.0298±0.02940.74±1.181.0417±0.04921.06±1.355.08±5.964.76±5.72 UniGradICON 6 92.78±10.6482.45±9.1378.36±23.7180.64±16.851.0105±0.04780.24±0.181.0196±0.06650.42±0.368.12±9.866.82±9.74 CAP-DIR (Clin.)97.92±2.1891.86±3.9494.88±4.2690.74±4.121.0041±0.00690.08±0.071.0076±0.01320.16±0.126.94±11.257.13±11.18 CAP-DIR (Clin.+Hist.) 98.61±1.0494.42±2.2696.92±2.5893.36±2.721.0013±0.00370.04±0.041.0038±0.00810.08±0.076.59±11.036.54±10.10 Best and second-best results are shown in bold and underline, respectively (displacement columns are descriptive, not ranked). 2 22/27 Supplementary Table 2. Per-site quantitative comparison of all registration methods on the internal test set. Results are reported for all eight methods across seven anatomical sites (586 patients, 2,122 registration cases from Mayo AZ and Mayo MN). One registration case corresponds to one source–target CT pair; clinical context and available patient history do not alter the case count. Values are reported as mean± standard deviation. All metrics follow the same definition as Supplementary Table 1: NCC and SSIM (%) for image similarity; mean Jacobian determinant (→1) and folding ratio (%↓) for deformation regularity, each computed at Body and OAR levels. CAP-DIR (Clin.) uses clinical context, and CAP-DIR (Clin.+Hist.) additionally uses patient history. Bold = best per metric and site; underline = second best. SiteRegistration casesMethod Image SimilarityDeformation Regularity BodyOARBodyOAR NCC(%)↑SSIM(%)↑NCC(%)↑SSIM(%)↑Jac→1Fold%↓Jac→1Fold%↓ Other16 VoxelMorph 1 92.29±8.7182.23±9.0866.95±26.3677.81±13.790.99±0.010.22±0.231.02±0.050.67±0.36 TransMorph 2 95.21±3.5887.82±5.7876.88±18.7385.36±8.651.06±0.041.88±1.791.08±0.062.39±1.97 CorrMLP 3 94.49±5.3988.08±6.2978.83±17.9886.10±8.581.04±0.020.96±1.381.06±0.061.29±1.56 DiffuseMorph 85.65±11.1574.89±9.5457.76±24.0170.60±13.341.00±0.010.07±0.081.00±0.010.18±0.19 MambaMorph 5 94.35±6.0185.46±7.6173.92±20.8782.73±10.181.04±0.031.06±1.241.05±0.051.43±1.41 UniGradICON 6 87.52±9.7777.94±8.3965.46±20.7771.47±14.761.03±0.050.32±0.181.04±0.070.57±0.35 CAP-DIR (Clin.)97.75±2.1691.55±3.8894.41±3.6790.34±3.801.00±0.010.09±0.061.01±0.010.17±0.10 CAP-DIR (Clin.+Hist.)98.51±0.9694.20±2.0996.64±2.2493.09±2.501.00±0.000.04±0.041.00±0.010.09±0.06 Breast161 VoxelMorph 1 97.61±6.9687.78±7.2584.92±21.0587.22±11.010.99±0.010.10±0.191.00±0.040.49±0.29 TransMorph 2 97.70±2.8691.83±4.6291.45±14.9592.09±6.911.04±0.030.85±1.431.06±0.051.25±1.57 CorrMLP 3 97.78±4.3091.93±5.0291.09±14.3591.95±6.851.03±0.020.26±1.101.04±0.050.50±1.24 DiffuseMorph 95.15±8.9083.02±7.6280.89±19.1783.45±10.651.00±0.010.01±0.061.00±0.010.04±0.15 MambaMorph 5 98.28±4.8090.43±6.0889.14±16.6690.16±8.131.02±0.030.39±0.991.03±0.040.66±1.13 UniGradICON 6 98.40±7.8187.27±6.7092.14±16.5890.43±11.790.99±0.040.16±0.141.00±0.050.26±0.28 CAP-DIR (Clin.)98.70±1.7293.31±3.1097.07±2.9392.60±3.031.00±0.010.04±0.051.00±0.010.10±0.08 CAP-DIR (Clin.+Hist.)99.07±0.7795.43±1.6798.22±1.7994.60±1.991.00±0.000.02±0.031.00±0.010.05±0.05 Brain160 VoxelMorph 1 93.62±8.2783.61±8.6371.44±25.0380.16±13.100.99±0.010.19±0.221.02±0.050.62±0.34 TransMorph 2 95.83±3.4088.82±5.4980.52±17.7887.04±8.221.05±0.041.62±1.701.08±0.062.11±1.87 CorrMLP 3 95.31±5.1289.04±5.9781.90±17.0787.56±8.151.04±0.020.78±1.311.05±0.051.10±1.48 DiffuseMorph88.02±10.5976.92±9.0663.54±22.8073.81±12.671.00±0.010.05±0.081.00±0.010.14±0.18 MambaMorph 5 95.33±5.7186.70±7.2377.73±19.8284.59±9.671.03±0.030.90±1.171.05±0.051.24±1.34 UniGradICON 6 90.24±9.2880.27±7.9772.13±19.7276.21±14.011.02±0.050.28±0.171.03±0.060.49±0.33 CAP-DIR (Clin.)97.17±2.0590.47±3.6992.78±3.4888.96±3.611.01±0.010.12±0.061.01±0.010.22±0.10 CAP-DIR (Clin.+Hist.)98.17±0.9293.45±1.9995.67±2.1392.17±2.371.00±0.000.06±0.041.01±0.010.11±0.06 Abdomen195 VoxelMorph 1 91.62±8.9381.53±9.3164.70±27.0376.64±14.140.99±0.010.23±0.241.03±0.050.69±0.37 TransMorph 2 94.90±3.6787.32±5.9275.06±19.2084.51±8.871.06±0.042.01±1.841.09±0.062.53±2.02 CorrMLP 3 94.08±5.5387.60±6.4577.30±18.4385.37±8.801.04±0.021.05±1.411.06±0.061.39±1.60 DiffuseMorph84.46±11.4373.88±9.7854.87±24.6168.99±13.681.00±0.010.07±0.081.00±0.010.19±0.20 MambaMorph 5 93.86±6.1784.84±7.8172.02±21.3981.80±10.441.04±0.031.15±1.271.05±0.051.53±1.45 UniGradICON 6 86.16±10.0276.77±8.6062.13±21.2969.10±15.131.03±0.050.34±0.181.05±0.070.61±0.36 CAP-DIR (Clin.)96.80±2.2189.79±3.9891.74±3.7688.08±3.891.01±0.010.13±0.061.01±0.010.25±0.11 CAP-DIR (Clin.+Hist.)97.95±0.9992.98±2.1595.05±2.3091.58±2.561.00±0.000.06±0.041.01±0.010.13±0.06 Pelvis421 VoxelMorph 1 90.62±9.2680.49±9.6561.33±28.0274.87±14.660.98±0.010.25±0.251.03±0.060.73±0.38 TransMorph 2 94.43±3.8186.56±6.1472.32±19.9183.25±9.201.06±0.042.20±1.901.09±0.072.74±2.09 CorrMLP 3 93.46±5.7386.88±6.6975.00±19.1184.27±9.121.04±0.021.18±1.471.06±0.061.54±1.66 DiffuseMorph82.68±11.8572.35±10.1450.53±25.5266.58±14.181.00±0.010.09±0.091.01±0.010.22±0.21 MambaMorph 5 93.13±6.3983.91±8.0969.17±22.1880.41±10.821.04±0.031.27±1.311.06±0.061.67±1.50 UniGradICON 6 84.12±10.3975.02±8.9257.13±22.0765.55±15.691.04±0.050.36±0.191.05±0.070.67±0.37 CAP-DIR (Clin.)96.53±2.2989.30±4.1391.00±3.9087.45±4.041.01±0.010.15±0.061.02±0.010.27±0.11 CAP-DIR (Clin.+Hist.)97.79±1.0292.64±2.2294.61±2.3891.17±2.651.00±0.000.07±0.041.01±0.010.14±0.07 Head & Neck669 VoxelMorph 1 95.61±7.6185.70±7.9478.18±23.0483.69±12.060.99±0.010.14±0.201.01±0.050.55±0.31 TransMorph 2 96.76±3.1390.33±5.0585.98±16.3789.56±7.561.05±0.031.23±1.571.07±0.061.68±1.72 CorrMLP 3 96.55±4.7190.48±5.5086.50±15.7189.76±7.501.03±0.020.52±1.211.04±0.050.80±1.36 DiffuseMorph91.58±9.7579.97±8.3472.21±20.9978.63±11.661.00±0.010.03±0.071.00±0.010.09±0.17 MambaMorph 5 96.81±5.2688.56±6.6683.43±18.2487.37±8.901.03±0.030.64±1.081.04±0.050.95±1.24 UniGradICON 6 94.32±8.5483.77±7.3382.14±18.1583.32±12.901.01±0.040.22±0.151.01±0.060.38±0.31 CAP-DIR (Clin.)98.33±1.8992.62±3.3996.03±3.2191.72±3.321.00±0.010.06±0.051.01±0.010.13±0.09 CAP-DIR (Clin.+Hist.)98.85±0.8494.95±1.8397.61±1.9694.01±2.181.00±0.000.03±0.031.00±0.010.06±0.06 Lung500 VoxelMorph 1 98.28±6.7488.47±7.0287.17±20.3988.39±10.670.99±0.010.08±0.181.00±0.040.46±0.28 TransMorph 2 98.01±2.7792.34±4.4793.27±14.4892.93±6.691.04±0.030.72±1.381.05±0.051.11±1.52 CorrMLP 3 98.19±4.1792.41±4.8692.62±13.9092.68±6.641.03±0.020.17±1.071.03±0.040.40±1.21 DiffuseMorph96.33±8.6284.03±7.3883.78±18.5785.06±10.321.00±0.010.00±0.061.00±0.010.02±0.15 MambaMorph 5 98.77±4.6591.05±5.8991.04±16.1491.08±7.871.02±0.020.31±0.961.03±0.040.57±1.09 UniGradICON 6 99.76±7.5688.44±6.4995.48±16.0692.80±11.410.99±0.040.14±0.140.99±0.050.22±0.27 CAP-DIR (Clin.)98.97±1.6793.79±3.0097.81±2.8493.23±2.941.00±0.010.03±0.051.00±0.010.08±0.08 CAP-DIR (Clin.+Hist.)99.23±0.7595.77±1.6298.67±1.7495.02±1.931.00±0.000.02±0.031.00±0.010.04±0.05 3 23/27 Supplementary Table 3. Registration performance across the seven held-out external cohorts. NCC and SSIM values are reported as mean± s.d. percentages across registration cases for the unregistered images, six comparison methods and the two CAP-DIR variants. CAP-DIR (Clin.) uses clinical context, whereas CAP-DIR (Clin.+Hist.) additionally uses available preceding patient history. All cohorts were excluded from model development and evaluated without retraining. Bold indicates the highest value within each cohort and metric. CohortCasesMetricInitialDemonsVoxelMorphTransMorphCorrMLPMambaMorphUniGradICONCAP-DIR (Clin.)CAP-DIR (Clin.+Hist.) NLST122Body NCC81.24±10.0485.48±7.9886.26±6.0587.64±2.5588.12±3.4187.38±4.3684.26±7.7589.18±1.6290.36±0.77 Body SSIM64.26±6.9669.48±6.6370.42±6.1672.85±3.7273.68±3.9871.94±5.3268.12±6.3076.42±2.7978.76±1.65 OAR NCC74.48±21.4981.56±17.0682.45±16.9787.26±12.5588.42±11.9486.18±13.5080.25±16.5593.46±3.0295.81±1.70 OAR SSIM61.84±10.9468.80±9.2670.20±9.9475.80±6.4377.60±5.6674.20±7.3767.40±11.3784.20±3.0086.80±1.87 LongCT-Lung353Body NCC66.28±11.7570.84±9.5371.62±6.6774.26±3.1175.14±4.6373.68±5.1070.12±8.8577.16±1.8578.79±0.91 Body SSIM62.64±9.1767.58±7.5368.64±7.7671.42±4.6572.48±4.8170.46±6.2966.25±8.0174.68±3.4677.00±1.90 OAR NCC64.86±23.2772.14±20.1673.28±20.7079.26±15.5480.65±15.2377.84±16.5570.42±21.1886.62±3.7089.45±2.20 OAR SSIM60.16±13.7066.80±10.8568.40±12.2373.60±6.9575.40±6.7472.00±9.0765.20±13.7479.20±3.6481.80±2.39 RIDER54Body NCC84.62±7.8188.96±6.7589.74±5.1691.28±2.1491.85±3.2690.94±3.7287.56±6.7992.72±1.3793.65±0.65 Body SSIM68.12±5.9372.84±5.8773.95±5.3376.26±3.2277.12±3.6775.24±4.6871.36±5.2279.72±2.3181.95±1.39 OAR NCC78.84±17.3085.28±14.0786.35±14.9790.82±11.8891.96±11.0489.54±11.9284.12±13.2395.58±2.5297.76±1.44 OAR SSIM64.85±9.2771.80±7.6973.40±8.6178.60±4.9880.40±5.1276.80±6.0470.00±10.0186.00±2.5888.40±1.59 Pancreas-CT80Body NCC47.26±11.7653.84±9.2255.12±6.4058.68±2.7860.15±4.2857.46±4.7952.28±8.3962.04±1.7263.72±0.82 Body SSIM19.85±8.1723.68±6.8524.42±6.7127.16±4.4228.45±4.7126.38±6.0722.64±7.4029.68±3.0631.26±1.70 OAR NCC48.65±23.5257.68±19.3858.92±20.1066.24±15.8668.16±14.0364.15±16.4154.86±18.5978.84±3.6482.48±2.02 OAR SSIM25.14±12.2231.80±10.0433.40±10.9740.20±7.0842.80±6.8237.80±8.1329.80±13.5750.80±3.3554.60±2.24 4D-Lung720Body NCC95.16±3.6697.24±2.7597.68±2.1898.46±0.8898.72±1.2898.28±1.4496.45±2.9999.08±0.5499.26±0.24 Body SSIM90.24±2.9892.56±2.6993.28±2.4994.86±1.5695.32±1.6194.48±2.1691.85±2.7196.14±1.0896.82±0.59 OAR NCC93.86±8.5196.64±6.3697.02±7.0898.26±4.2998.64±4.1597.98±5.0895.84±6.2199.18±0.9699.46±0.63 OAR SSIM87.26±4.7389.80±3.7290.60±3.7792.40±2.5293.00±2.3791.80±2.8089.00±4.9793.60±1.2094.20±0.79 NYPC21Body NCC84.28±10.5989.84±8.2190.76±5.9993.42±2.6794.28±3.9792.86±4.5088.42±7.6996.05±1.5397.12±0.73 Body SSIM74.86±8.2780.24±7.1881.18±6.5085.26±4.4586.42±4.5684.16±5.8178.84±6.8889.64±2.8691.86±1.60 OAR NCC70.42±22.8578.26±17.9679.45±20.0385.28±14.1786.62±12.4683.85±16.3776.14±17.6090.26±3.1292.84±2.07 OAR SSIM68.00±11.2574.00±9.9575.60±10.7480.80±6.5282.60±6.1079.20±7.2772.00±12.0385.60±3.1987.80±2.06 MD Anderson23Body NCC83.15±10.4188.76±8.9589.62±6.9192.54±2.9193.45±4.0891.88±4.4687.26±8.2195.32±1.5996.58±0.74 Body SSIM72.42±8.2978.16±7.5579.24±7.2683.58±4.1484.86±4.5682.25±5.8976.68±6.7388.14±3.1590.42±1.73 OAR NCC69.25±23.5677.18±17.4378.36±20.8884.42±15.3485.85±12.8282.74±16.4774.86±18.5589.42±3.3392.15±2.11 OAR SSIM66.60±12.9272.80±10.5374.40±10.8579.40±6.5081.20±6.7177.80±7.7770.60±13.2384.20±3.4486.60±2.03 4 24/27 Supplementary Table 4. Blinded clinical assessment of propagated contours. Twenty longitudinal planning-CT cases were independently assessed by two radiation oncology reviewers, yielding 40 reader–case evaluations per method. Clinical acceptability was defined as a rating of 3 or 4 (no more than minor edits). Editing-time estimates were averaged between reviewers within each case and are reported as the median (interquartile range) across 20 cases. Confidence intervals for acceptability differences were obtained by patient-level bootstrap resampling. Inter-reviewer agreement was assessed using linearly weighted Cohen’sκ ; the shared value in each structure block spans the two method columns. StructureEndpointCorrMLPCAP-DIRCAP-DIR benefit CTVClinically acceptable, n/N (%)18/40 (45.0)33/40 (82.5)+37.5 percentage points (95% CI, 15.0–60.0) Rating, median (IQR)2.0 (1.0–3.0)3.0 (3.0–3.0)— Estimated editing time, median (IQR), min11.3 (7.5–25.6)8.8 (7.5–10.0)−2.5 min (22.2% reduction) Inter-reviewer agreement, weightedκ0.58— OARClinically acceptable, n/N (%)9/40 (22.5)25/40 (62.5)+40.0 percentage points (95% CI, 17.5–60.0) Rating, median (IQR)2.0 (1.0–2.0)3.0 (2.0–3.0)— Estimated editing time, median (IQR), min26.3 (14.4–45.6)15.0 (12.5–18.1)−11.3 min (42.9% reduction) Inter-reviewer agreement, weightedκ0.62— 5 25/27 Supplementary Table 5. Site-specific dose-mapping results in the controlled synthetic benchmark. Seventy prescribed-deformation cases were evaluated across five anatomical sites. Global, structure-region and gradient-boundary values are mean± s.d. absolute errors expressed as a percentage of the case-specific dose scale. DVH endpoint error is the mean±s.d. absolute error across OAR and CTVD mean andD max , normalised by prescription dose. Reduction rows compare CAP-DIR with CorrMLP; for the three voxel-level metrics, reductions were calculated from site-level mean errors, whereas the DVH value is the mean endpoint-wise relative reduction, consistent with Fig. 4d. Lower error is better. SiteCasesMethodGlobal MAE (%)OAR/CTV-region MAE (%)Gradient-boundary MAE (%)DVH endpoint error (%) Brain/Head12No reg.1.85±0.632.64±0.9411.45±3.922.17±0.93 Demons1.44±0.462.07±0.728.93±2.871.58±0.62 CorrMLP1.10±0.351.59±0.496.81±1.921.20±0.39 CAP-DIR0.73±0.211.04±0.304.47±1.280.83±0.26 CAP-DIR reduction vs CorrMLP (%)33.634.534.330.3 Head & Neck26No reg.2.05±0.642.90±1.0012.70±4.121.25±0.45 Demons1.58±0.462.25±0.709.81±2.970.94±0.33 CorrMLP1.13±0.341.58±0.496.88±1.730.75±0.25 CAP-DIR0.72±0.181.03±0.284.50±1.100.46±0.14 CAP-DIR reduction vs CorrMLP (%)35.934.734.638.1 Breast4No reg.1.18±0.441.64±0.707.21±2.861.37±0.58 Demons0.94±0.361.35±0.505.72±2.101.05±0.41 CorrMLP0.74±0.261.04±0.404.64±1.570.84±0.31 CAP-DIR0.53±0.160.77±0.243.30±0.970.56±0.20 CAP-DIR reduction vs CorrMLP (%)27.726.628.932.4 Abdomen12No reg.2.01±0.702.88±1.1412.38±4.441.99±0.85 Demons1.60±0.602.25±0.879.82±3.181.47±0.58 CorrMLP1.23±0.431.75±0.617.59±2.341.17±0.43 CAP-DIR0.80±0.241.12±0.364.83±1.340.74±0.27 CAP-DIR reduction vs CorrMLP (%)35.136.036.337.2 Pelvis16No reg.1.50±0.532.12±0.789.11±3.141.58±0.60 Demons1.18±0.361.65±0.587.17±2.291.13±0.39 CorrMLP1.03±0.291.46±0.456.37±1.740.82±0.27 CAP-DIR0.69±0.200.98±0.294.21±1.120.57±0.17 CAP-DIR reduction vs CorrMLP (%)33.432.733.930.5 6 26/27 References 1.Balakrishnan, G., Zhao, A., Sabuncu, M. R., Guttag, J. & Dalca, A. V. VoxelMorph: A learning framework for deformable medical image registration. IEEE Transactions on Med. Imaging 38, 1788–1800, DOI: 10.1109/TMI.2019.2897538 (2019). 2.Chen, J. et al. TransMorph: Transformer for unsupervised medical image registration. Med. Image Analysis 82, 102615, DOI: 10.1016/j.media.2022.102615 (2022). 3.Meng, M., Feng, D., Bi, L. & Kim, J. Correlation-aware coarse-to-fine MLPs for deformable medical image registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9645–9654, DOI: 10.1109/CVPR52733.2024.00921 (2024). 4.Kim, B., Han, I. & Ye, J. C. DiffuseMorph: Unsupervised deformable image registration using diffusion model. In Proceedings of the European Conference on Computer Vision (ECCV), vol. 13662 of Lecture Notes in Computer Science, 347–364, DOI: 10.1007/978-3-031-20086-1_20 (Springer, Cham, 2022). 5.Guo, T. et al. MambaMorph: a Mamba-based framework for medical MR-CT deformable registration. arXiv preprint arXiv:2401.13934 DOI: 10.48550/arXiv.2401.13934 (2024). 6.Tian, L. et al. uniGradICON: A foundation model for medical image registration. In Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, vol. 15002 of Lecture Notes in Computer Science, 749–760, DOI: 10.1007/978-3-031-72069-7_70 (Springer, Cham, 2024). 7 27/27