PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2610.18926v1eess.IVcs.AI

TL;DR

Adaptive radiotherapy needs reliable correspondence across serial images to assess anatomical and dosimetric changes, but common pairwise registration does not use radiotherapy-specific context or a patient’s preceding trajectory. CAP-DIR combines source-side clinical information with a memory of prior registrations. It improved OAR-focused registration across internal and external cohorts and showed gains in contour usability and controlled dose mapping.

  • Problem

    Pairwise DIR commonly treats images independently and does not explicitly use radiotherapy-specific clinical context or the patient’s preceding anatomical trajectory.

    Source: pp. 1–2

  • Method

    CAP-DIR conditions registration on source-side contours, dose and clinical text, and incorporates temporally ordered features from prior registrations for that patient.

    Source: p. 2

  • Results

    Across internal and seven external cohorts, CAP-DIR’s largest image-similarity gains were in OAR regions and extended to contour propagation, blinded review and controlled dose mapping.

    Source: II Discussion · pp. 4–5

  • Takeaways & Limitations

    The findings support clinically informed, patient-adaptive registration as a means of providing more reliable starting contours and longitudinal dose estimates for adaptive assessment.

    Source: III Conclusion · p. 5

    Model development used data from two Mayo Clinic system centres, and generalisability beyond longitudinal CT-to-CT registration remains to be established.

    Source: II Discussion · p. 5

Abstract

from arXiv · show

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.

I.1 Study design and cohorts

CAP-DIR uses source-side clinical information and prior registrations to estimate each patient’s planning-to-treatment-state deformation without using future images or target-side contours. The evaluation used fixed longitudinal pairs across internal and held-out external cohorts.

I.1 Study design and cohorts

  • CAP-DIR conditions registration on source-side contours, dose and structured clinical text, then adds features from earlier registrations for the same patient.Only registrations preceding the current target image enter the longitudinal memory; target-side contours are used only for evaluation.Source: §I.1 · p. 2
  • The dataset comprised 3,473 patients, 5,005 plans, 16,361 CT examinations and 11,986 source–target image pairs.Cohorts covered seven anatomical sites and more than 20 disease categories.Source: §I.1 · p. 2
  • Model development used patient-level partitions of two Mayo Clinic cohorts, while MD Anderson, New York Proton Center and five public benchmarks were held out for external evaluation.The internal partitions included 2,052 training, 293 validation and 586 testing patients; fixed pairs were used for evaluation.Source: §I.1 · p. 2

I.2 Internal and external registration performance

CAP-DIR achieved stronger internal and external image-similarity results than the comparators, with its largest internal advantage in OAR regions. The OAR gains were accompanied by low folding and persisted across seven external cohorts.

I.2 Internal and external registration performance

  • 96.92% OAR-region NCC was achieved by CAP-DIR, versus 84.76% for CorrMLP on the internal test set.CAP-DIR also had the highest values across all four reported image-similarity measures.Source: §I.2 · p. 2
  • The OAR folding ratio was 0.08% for CAP-DIR and 0.91% for CorrMLP, alongside an OAR-region Jacobian deviation of 0.0038 for CAP-DIR.Source: §I.2 · p. 3
  • Adding patient history reduced the body-to-OAR NCC gap from 3.04 to 1.69 percentage points, after clinical context had reduced it from 11.32 points.The regional advantage was observed across all seven anatomical sites.Source: §I.2 · p. 3
  • Across seven external cohorts, CAP-DIR had the highest values for all four reported image-similarity measures, with OAR-region NCC gains of 0.8–14.3 percentage points over each cohort’s strongest comparator.Absolute OAR-region NCC ranged from 82.5% in Pancreas-CT to 99.5% in 4D-Lung.Source: §I.2 · p. 3

I.3 Contour propagation and blinded review

CAP-DIR improved propagated-contour agreement and reduced residual contour error relative to registration comparators. In blinded review, more propagated contours were rated clinically acceptable and estimated editing time was lower.

I.3 Contour propagation and blinded review

  • OAR DSC was 92.07% with CAP-DIR versus 82.28% with CorrMLP, while CTV DSC was 87.14% versus 73.32%.Residual symmetric-difference volume decreased from 9.66 to 2.06 ml on the common 1-mm grid.Source: §I.3 · p. 3
  • Clinically acceptable contours increased from 45.0% to 82.5% for CTVs and from 22.5% to 62.5% for OARs in the blinded review.Median estimated editing time decreased from 11.3 to 8.8 min for CTVs and from 26.3 to 15.0 min for OARs.Source: §I.3 · p. 3
  • The largest absolute increase in clinical acceptability occurred for OARs, which often comprise several propagated structures and create a greater editing burden.Objective contour analysis and reviewer assessments agreed that anatomical improvement extended to clinical usability.Source: §I.3 · p. 3

I.4 Controlled dose-mapping benchmark

In a controlled synthetic benchmark with known deformations, CAP-DIR reduced global dose-mapping error relative to the strongest comparator. Structure-level dose endpoint errors were also lower than with CorrMLP.

I.4 Controlled dose-mapping benchmark

  • 0.72% mean absolute error of the dose scale was achieved by CAP-DIR, versus 1.10% for CorrMLP, reducing global dose error by 34.5%.The benchmark used prescribed treatment-course deformations to compare recovered accumulated dose with ground truth.Source: §I.4 · pp. 3–4
  • Average error across OAR and CTV mean- and maximum-dose endpoints decreased from 0.97% with CorrMLP to 0.64% with CAP-DIR.The controlled known-deformation benchmark does not constitute prospective validation of clinical delivered-dose accumulation.Source: §I.4 · p. 4

I.5 Contribution of clinical context

Clinical context improved registration, with the largest gains in OAR regions; combining structured text with contours and dose produced the strongest results.

I.5 Contribution of clinical context

  • 94.9% OAR-region NCC with combined context exceeded 84.4% with image-only registration.The combined model also achieved 97.9% body-region NCC, versus 95.8% for image-only registration.Source: §I.5 · p. 4
  • Combining text, contours and dose increased OAR-region NCC by 10.5 percentage points and narrowed the body-to-OAR gap from 11.4 to 3.0 percentage points.Text alone increased OAR-region NCC by 5.7 points, while contours and dose increased it by 8.0 points.Source: §I.5 · p. 4
  • Combined context improved OAR-region NCC across lung, breast, abdomen and pelvis, with gains of 5.4, 6.3, 14.9 and 16.5 percentage points, respectively.The largest gains occurred in abdominal and pelvic cases.Source: §I.5 · p. 4

I.6 Contribution of patient history

Registration history improved OAR alignment beyond clinical context alone across every evaluated history length, with benefits varying by anatomy and history length.

I.6 Contribution of patient history

  • 97.1% to 96.0% OAR-region NCC with history-conditioned registration exceeded 95.5% to 92.5% with clinical context alone across the same history strata.The history-conditioned model improved OAR-region NCC at every evaluated history length.Source: §I.6 · p. 4
  • History added 1.6 to 3.5 percentage points of OAR-region NCC, compared with 0.6 to 1.0 points in the body region.The OAR gains were larger than the body-region gains across reported history lengths.Source: §I.6 · p. 4
  • Abdominal cases had the largest reported history-related OAR gains, ranging from 2.2 to 5.2 percentage points across strata.Brain and breast gains were 0.9–2.0 points, while shorter-history head-and-neck and pelvic gains were 2.4–2.7 points.Source: §I.6 · p. 4

II Discussion

CAP-DIR’s gains extended across external cohorts, contour assessment and dose mapping. The discussion attributes its clinically focused performance to clinical context and patient history, while noting limits on generalizability and validation.

II Discussion

  • CAP-DIR maintained a relative advantage across seven external cohorts, and gains extended beyond image similarity to contour propagation, blinded review and controlled dose mapping.The blinded review found more clinically acceptable propagated contours with less editing; the dose benchmark found lower errors in accumulated-dose maps and structure-level summaries.Source: II Discussion · pp. 4–5
  • OAR-focused gains matter because whole-image similarity can be dominated by large, stable regions, and CAP-DIR is intended to provide a starting point for clinician review, not replace it.The discussion reports improved acceptability for both CTVs and OARs, with particularly marked improvements for OAR acceptability and editing requirements.Source: II Discussion · p. 5
  • Contours and dose provide treatment-relevant spatial context, text supplies semantic context, and prior registrations add patient-specific history without target-side contours or inference-time fine-tuning.The discussion describes planning contours, dose and metadata as reusable from treatment preparation, with memory updated as treatment images become available.Source: II Discussion · p. 5
  • The dose benchmark used prescribed deformations as a controlled reference because true patient deformation is unavailable for voxelwise validation.The external cohorts and public benchmarks varied in anatomy, acquisition and follow-up; performance remained most appropriately interpreted within each cohort.Source: II Discussion · p. 5
  • Generalizability remains bounded by development at two Mayo Clinic system centres and evaluation limited to longitudinal CT-to-CT registration; clinician-approved contours also vary between observers.The discussion calls for broader prospective evaluation across institutions and practice settings, and evaluation with other modalities and cross-modality registration.Source: II Discussion · p. 5
  • Memory uses features from preceding Stage 1 registrations, so errors in those estimates could propagate into later registrations; robustness to this remains unevaluated.This is a model-specific dependency of the longitudinal design.Source: II Discussion · p. 5

III Conclusion

The conclusion presents clinically informed, patient-adaptive registration as a means of improving treatment-relevant alignment across the treatment course. Its reported gains extend to contour propagation, clinical assessment and dose mapping.

III Conclusion

  • Integrating radiotherapy-specific clinical context with longitudinal patient history can benefit DIR for adaptive radiotherapy.The conclusion frames CAP-DIR as a clinically informed and patient-adaptive formulation.Source: III Conclusion · p. 5
  • Across internal testing and seven held-out external cohorts, CAP-DIR preferentially improved treatment-relevant OAR registration while maintaining plausible deformations.Source: III Conclusion · p. 5
  • The reported gains extended to contour propagation, blinded clinical assessment and controlled dose mapping, supporting more reliable starting contours and longitudinal dose estimates for adaptive assessment.Source: III Conclusion · p. 5

IV.1 Ethics

The multi-site retrospective study received single-IRB approval, with consent waived because of its retrospective design.

IV.1 Ethics

  • The Mayo Clinic Institutional Review Board served as the single IRB of Record for participating sites, and the study followed the Declaration of Helsinki.Source: IV.1 Ethics · p. 5
  • The IRB waived study-specific written informed consent because of the retrospective design.Source: IV.1 Ethics · p. 6
  • Data use followed institutional arrangements and the original access and data-use requirements for public datasets.Source: IV.1 Ethics · p. 6

IV.2 Study datasets and preprocessing

This section describes the datasets and preprocessing used to develop and evaluate CAP-DIR, including longitudinal radiotherapy CT pairs, source-derived clinical context, patient history, data splits, and matched evaluation procedures. It also covers supplementary analyses of dose mapping and contour propagation, including how history-related gains varied with available prior observations.

IV.2 Study datasets and preprocessing

  • CTs were harmonised to common physical space, resampled to 1×1×1 mm3, cropped or padded to 512×512×256, and intensity-normalised; contours and dose maps received corresponding spatial transformations.Clinical context combined structured text, source contours and dose; missing public-dataset inputs were derived from source-side information, while target annotations were reserved for evaluation.Source: IV.2 Study datasets and preprocessing · p. 6

IV.3 Image-pair construction and dataset partitioning

  • The dataset comprised 3,473 patients, 5,005 treatment plans, 16,361 CT examinations and 11,986 image pairs, constructed from eligible longitudinal scans within patients and treatment courses.Pairing used planning-to-verification scans, later plans, or baseline-to-later benchmark examinations, depending on cohort structure.Source: IV.3 Image-pair construction and… · p. 6
  • Mayo AZ and Mayo MN patients were split 70%/10%/20% for training, validation and internal testing; MD Anderson, NYPC and five public datasets were held out for external validation.All scans from each patient stayed in one split, and Stage 2 histories included only temporally prior plans or fractions.Source: IV.3 Image-pair construction and… · p. 6

IV.4 CAP-DIR framework

  • CAP-DIR uses a shared image-registration backbone conditioned first on semantic text and spatial contour-dose context, then on a memory of prior images, deformations and registration quality.Stage 2 summarizes temporally ordered history with a GRU; no manual target contours are used as model inputs.Source: IV.4 CAP-DIR framework · pp. 7–8
  • The registration objective weights local NCC in foreground body and OAR regions and applies diffusion regularisation, with λOAR = 1.0 and λreg = 0.2 unless specified otherwise.The OAR mask is derived from source contours warped into target space, not manual target contours.Source: IV.4 CAP-DIR framework · p. 9

IV.5 Implementation and comparison methods

  • Comparisons included learning-based and conventional DIR methods; ablations contrasted image-only, clinical-context and full clinical-context-plus-history models using the same image pairs and patient-level splits.CAP-DIR training used sequential Stage 1 and Stage 2 training, with validation-selected checkpoints for each stage.Source: IV.5 Implementation and comparison… · p. 9

IV.6 Blinded clinical contour review

  • A blinded review assessed unedited CAP-DIR and CorrMLP contours against clinician-approved references in 20 longitudinal cases, with two reviewers rating CTVs and OARs and estimating editing time.Acceptability for analysis meant a rating of 3 or 4; the review used routine clinical software and contours resampled to a common grid.Source: IV.6 Blinded clinical contour review · p. 9, IV.7 Evaluation metrics · p. 10

IV.7 Evaluation metrics; IV.8 Statistical analysis

  • Evaluation covered image similarity, deformation plausibility, contour overlap and residual volume, clinical acceptability and editing time, and dose MAE and DVH endpoint error.Analyses used matched evaluation units, paired Wilcoxon tests when assessing significance, and patient-level bootstrap confidence intervals for clinical-acceptability differences.Source: IV.7 Evaluation metrics · pp. 9–10, IV.8 Statistical analysis · p. 10

Figures

  • The framework encodes shared-weight source and target CTs, then uses coarse-to-fine transformer decoding; clinical context conditions Stage 1 and patient history conditions Stage 2.Clinical text becomes BioMedBERT embeddings and FiLM parameters, while contours and dose maps form spatial anatomy and risk maps; a GRU encodes prior image features, deformation statistics and quality memory.Source: Figure 7 · p. 19

Tables

  • The dataset split overview reports patient, treatment-plan, CT-examination and image-pair counts across training, validation, internal-test and external-validation cohorts.Source: Table 1 · p. 20

Supplementary Notes

  • Residual symmetric-difference volume fell from 9.66 to 2.06 ml with CAP-DIR, a 79% reduction versus direct transfer and 55% versus CorrMLP.The analysis used longitudinal planning CTs: 276 internal source–target pairs, plus 23 from MD Anderson and 21 from the New York Proton Center.Source: Supplementary Notes · p. 21
  • The OAR-region gain from Stage 2 increased from 1.6 percentage points with one prior observation to 3.5 points with 17; body-region gains ranged from 0.6 to 1.0 points.Abdominal cases showed the largest observed site-specific increase.Source: Supplementary Notes · p. 21

Supplementary Tables

  • On the internal test set of 2,122 registration cases from 586 patients, CAP-DIR (Clin.+Hist.) achieved body-region SSIM of 96.92±2.58% (mean ± s.d.), with both variants evaluated on the full set.For cases without prior fraction history, CAP-DIR (Clin.+Hist.) falls back to Stage 1 behaviour.Source: Supplementary Tables · p. 22, Table 1 · p. 22

Supplementary Notes

  • 34.5% lower whole-volume dose mean absolute error and 33.3% lower mean error across OAR and CTV dose endpoints were achieved relative to CorrMLP.Across anatomical sites, reductions relative to CorrMLP were 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.Source: Supplementary Notes · p. 21, Table 5 · p. 26

Related papers

Loading 2610.18926v1…