Source-linked AI summary
An Uncertainty-Guided Digital Twin Framework for Online Adaptive Proton Therapy in Head and Neck Cancer: A Feasibility Study
Yizhou Wu, Ryan J. Sanford, Huiqiao Xie, Jie Ding, Shupeng Chen, Tung-Ho Wu, Ping-Hsiu Wu, Justin Roper, Jun Zhou, Minglei Kang, Bill Stokes, Sibo Tian, David S. Yu, Xiaofeng Yang, Chih-Wei Chang
TL;DR
Head and neck proton therapy spans weeks of progressive anatomical change, while offline replanning takes time. This study forecasts treatment-day anatomy with an uncertainty-guided digital twin and tests online adaptive plans against physician-approved offline replans. After QACT reoptimization, target coverage and critical-organ tolerance were comparable to the offline baseline in this ten-patient feasibility study.
Problem
Head and neck anatomy changes over a six-to-seven-week proton therapy course, and offline replanning can take about a week.
Source: §1 · p. 2Method
UGDT transports prior patients’ longitudinal deformations onto a new patient’s planning CT, derives a direction-resolved CTV robust margin from contour uncertainty, and uses the QACT to guide online plan reoptimization.
Source: §1 · p. 3, §2.4 · p. 8Results
After QACT reoptimization, CTV-H D98% was 98.3 ± 0.3% versus 98.5 ± 0.4% for the offline replan, with spinal cord and brainstem below tolerance in every patient.
Source: §4 · p. 20Takeaways & Limitations
In this ten-patient feasibility study, UGDT generated online adaptive plans comparable in quality to physician-approved offline replans using anatomy forecast before treatment.
Source: §5 · p. 22The single-institution cohort included ten patients and one replanning QACT each, so results characterize feasibility rather than generalization or intervening fractions.
Source: §4 · p. 21
Abstract
from arXiv · showhide
Objective: Head and neck (HN) proton therapy spans six to seven weeks of anatomical change, while offline replanning takes about a week. We present an uncertainty-guided digital twin (UGDT) framework that forecasts treatment-day anatomy before treatment and evaluate whether it generates online adaptive proton therapy (APT) plans of clinical quality. Approach: A library of 302 longitudinal deformations from 88 previously treated HN patients was transported onto each new patient's treatment planning CT (TPCT) using two-step multi-atlas deformable image registration (DIR) built on a pretrained CT foundation model, generating about 284 predicted CTs (pdCTs) with contours per patient. Dispersion of propagated clinical target volume (CTV) contours defined a patient-specific robust margin. In ten patients, the quality assurance CT (QACT) triggering a replan represented treatment-day anatomy, and the physician-approved replan was the baseline. The pdCT most similar to the QACT (pdCT-H) and one from the lowest quartile (pdCT-L) were planned to within about 5% of baseline plan quality, forward-calculated on the QACT, and reoptimized to generate online APT plans. Main results: pdCT plans scored within -0.7% (pdCT-H) and -1.0% (pdCT-L) of baseline. Forward calculation on QACT reduced high-dose CTV D98% to 88.3% and 85.5%. After online reoptimization, D98% recovered to 98.3 +/- 0.3% and 98.2 +/- 0.3%, versus 98.5 +/- 0.4% at baseline. Spinal cord and brainstem doses remained below tolerance, and plan quality scores were within -1.1% (p = 0.19) and -1.7% (p = 0.01) of baseline. Significance: UGDT generated online APT plans comparable in quality to physician-approved offline replans using anatomy forecast before treatment, enabling a transition from reactive offline replanning toward anticipatory online adaptation.
1. Introduction
Head and neck anatomy changes progressively during proton therapy, while offline replanning can leave patients on a plan for anatomy they no longer have. This work proposes forecasting anatomy before treatment and evaluates whether that forecast can support clinically comparable online adaptive plans.
1. Introduction
- 59 of 88 patients (67%) required an offline replan, with a median 9-day delay from the triggering QACT to the replanned course.Fractions delivered during that interval use a plan optimized for anatomy the patient no longer has.Source: §1 · p. 2
- Existing online adaptation starts after treatment-day imaging, while prior anatomy-forecasting models had been evaluated in the image domain without showing clinically acceptable proton plans could be built from them.For head and neck plans, contouring and robust reoptimization are time-consuming steps on the treatment-session critical path.Source: §1 · p. 2
- The digital-twin concept forecasts possible patient anatomies before treatment and synchronizes the twin with daily imaging to support treatment-day decisions.Prior head and neck work transported anatomical changes and quantified contour uncertainty, but remained in the image domain.Source: §1 · pp. 2–3
- The proposed UGDT transports prior anatomical changes onto a new patient’s planning CT, derives a patient-specific CTV robust margin, and tests online plans against the physician-approved offline replan on the QACT.Initial predicted-CT plans were controlled to within about 5% of the baseline plan quality before QACT evaluation and reoptimization.Source: §1 · p. 3
2. Materials and methods
The methods build an offline ensemble of patient-specific predicted CTs and contour-derived margins, then use treatment-day imaging to select and reoptimize a plan. The study compares these online plans with physician-approved offline replans using shared planning objectives and a knowledge-based quality score.
2.1 Uncertainty-guided digital twin framework
- The offline stage builds an ensemble of predicted CTs and direction-resolved CTV margins before treatment; on treatment day, QACT similarity selects a plan for evaluation and reoptimization.The retrospective evaluation uses the QACT-triggered physician-approved replan as the clinical baseline and tests both the best-matching and a low-quartile ensemble member.Source: §2.1 · pp. 3–5
2.2 Patient data and clinical replanning context
- The deformation library contains 302 prior-session changes from 88 patients, while ten patients with clinical replans form the evaluation cohort.The QACT-triggered replan serves as the treatment-day reference and quality baseline; the median interval from triggering QACT to replanned-course start was 9 days.Source: §2.2 · pp. 5–6, Figure 2 · p. 5
2.3 Foundation model-based two-step multi-atlas deformable image registration
- Two-step transport first maps a prior patient's planning and treatment scans to the new patient's grid, then estimates that patient's longitudinal change in the new frame for warping the new patient's CT and contours.Transported displacement vectors are reoriented by the inverse Jacobian of the cross-patient map; omitting this term produced a median relative error of 36.3% in the measured fields.Source: §2.3 · p. 7
2.4 Patient-specific CTV robust margin from contour uncertainty
- The patient-specific CTV robust margin is one standard deviation of propagated contour displacement in each of six directions, used as setup uncertainty in robust optimization.The margin comes from the same ensemble as the predicted CTs and replaces a fixed isotropic margin with direction-resolved values.Source: §2.4 · p. 8
2.5 Image similarity metrics and candidate ranking
- Candidate CTs are ranked against QACT using an equally weighted image-only composite of normalized cross-correlation, structural similarity, and normalized mean absolute error.The ranking selects pdCT-H at rank one and pdCT-L as the highest-ranked member of the lowest quartile; it uses no contour-derived term.Source: §2.5 · pp. 8–9
2.6 Treatment planning
- Four plan sets are compared: the clinical QACT replan, plans optimized on pdCT-H and pdCT-L, and those pdCT plans reoptimized on QACT as online APT plans.A patient-specific objective template is reused, with the ensemble-derived margins and 3.5% range uncertainty; pdCT plans are controlled to within about 5% of baseline score before QACT evaluation.Source: §2.6 · p. 9
2.7 Knowledge-based plan quality evaluation and statistical analysis
- A knowledge-based score sums criterion-specific target and organ-at-risk dose scores, with higher totals indicating plans closer to intended quality; paired comparisons use a two-sided Wilcoxon signed-rank test.The score difference is calculated relative to the offline QACT replan baseline, and individual dose-volume endpoints are also reported.Source: §2.7 · pp. 9–10
3. Results
Across ten patients, pdCT plans were acceptable starting points, but forward calculation on treatment-day QACT showed target-coverage loss without adaptation. Online reoptimization restored coverage and approached the physician-approved offline replans.
3.1 Ranking of predicted CTs for planning image set selection
- pdCT-H Cimg ranged from 0.937 to 0.967 (mean 0.953), versus 0.915 to 0.945 (mean 0.930) for pdCT-L; all three similarity terms were lower for pdCT-L.The median composite difference between the selected candidates was 0.0217.Source: §3.1 · p. 10, Table 1 · p. 11
3.2 Patient-specific CTV robust margin
- Contour-derived margins ranged from 1.1 to 3.0 mm (mean 1.8 mm); mean margins were largest inferiorly at 2.1 mm and smallest superiorly at 1.5 mm.Margins differed more between patients than directions and were used as setup uncertainty in robust optimization.Source: §3.2 · pp. 11–12, Table 2 · p. 12, Figure 4 · p. 12
3.3 Initial treatment planning using pdCT
- Initial pdCT plans preserved target coverage on their own images and scored within -0.7% (pdCT-H) and -1.0% (pdCT-L) of the offline replan.The plans met the intended 5% score control and were clinically acceptable starting points for treatment-day evaluation.Source: §3.3 · p. 13, Table 3 · p. 14
3.4 Online dose evaluation of pdCT plans on the QACT
- On QACT without reoptimization, CTV-H D98% averaged 88.3% (pdCT-H) and 85.5% (pdCT-L), so forecast-anatomy plans required adaptation before delivery.Only four of twenty plans retained CTV-H D98% above 95%.Source: §3.4 · p. 15, Table 4 · p. 16
3.5 Online adaptive treatment planning
- Online reoptimization restored CTV-H D98% to 98.3 ± 0.3% (pdCT-H) and 98.2 ± 0.3% (pdCT-L), versus 98.5 ± 0.4% baseline; quality scores were within -1.1% and -1.7%.The pdCT-H score difference was not significant (p = 0.19), while the pdCT-L difference was significant (p = 0.01).Source: §3.5 · p. 17, Table 5 · p. 18
4. Discussion
In this ten-patient feasibility study, plans built on forecast anatomy were close to the physician-approved offline replans, and QACT reoptimization restored target coverage. Forecast matching was more informative for post-adaptation plan quality than for individual forward-calculated coverage; margin sufficiency and generalization remain unresolved.
4. Discussion
- Online plans initialized from pdCT-H restored CTV-H D98% to 98.3 ± 0.3% versus 98.5 ± 0.4% at baseline, with a nonsignificant -1.1% average plan-quality difference.Forecast plans scored within about 5% of baseline, and nine of ten pdCT-H plans were within ±2%.Source: §4 · p. 20
- Forward calculation reduced average CTV-H D98% by 10.6 percentage points for pdCT-H and 13.0 for pdCT-L, while QACT reoptimization restored coverage to at least 98.0% across twenty plans.Only four of twenty forward-calculated plans retained CTV-H D98% above 95%.Source: §4 · pp. 20–21
- Image similarity did not reliably order individual forward-calculated coverage loss, but pdCT-H-initialized plans scored closer to baseline in seven of ten patients.Only the pdCT-L set differed significantly from baseline after reoptimization (-1.7%, p = 0.01).Source: §4 · p. 21
- CTV robust margins ranged from 1.1 to 3.0 mm, reflecting patient-specific, direction-resolved contour spread; their sufficiency cannot be established against replans using different clinical margins.Plans using these margins met target coverage after reoptimization in all twenty cases.Source: §4 · p. 21
- The ten-patient, single-institution cohort with one replanning QACT per patient supports feasibility, not generalization or performance on intervening treatment days.Dose-volume endpoints also depended on physician-drawn QACT contours, whose variability was not assessed.Source: §4 · p. 21
5. Conclusions
The study developed a dose-domain UGDT framework for head and neck online adaptive proton therapy. In ten patients, plans initialized from the best-matching forecast approached the physician-approved offline replan in plan quality and matched target coverage and organ-at-risk tolerance after reoptimization.
5. Conclusions
- In ten patients, QACT-reoptimized plans initialized from the best-matching forecast matched the offline replan in target coverage and spinal cord and brainstem tolerance, with plan quality within -1.1%.The framework forecasts anatomy before the course, derives a direction-resolved CTV robust margin from contour uncertainty, and selects forecasts by image similarity.Source: §5 · p. 22
Ethical statement
The retrospective analysis received institutional review board approval, and informed consent was not required.
Ethical statement
- The retrospective analysis received institutional review board approval; informed consent was not required.The research followed the Declaration of Helsinki and local statutory requirements.Source: Ethical statement · p. 22