Synthetic-Breathing Predictive Model

• The paper presents a synthetic-breathing-based predictive model integrated in RayStation to assess interplay effects in free-breathing IMPT.
• It utilizes a library of 24 sinusoidal breathing traces combined with 4DCT phases and machine delivery log data for explicit temporal mapping.
• Validated against Monte Carlo FRED and in vivo measurements, the model shows dose deviations within 2%, underlining its clinical feasibility.

Searching arXiv for the primary paper and a closely related motion/interplay paper to ground the article in current literature. A synthetic-breathing-based predictive model, in the context of free-breathing intensity-modulated proton therapy, is a pre-treatment 4D dose evaluation framework that combines artificial respiratory traces, planning 4DCT data, and machine delivery timing to predict interplay effects and motion robustness before treatment starts. In the reported implementation, the framework was developed for lung and esophageal cancer patients treated with pencil beam scanning proton therapy under free-breathing conditions, integrated in RayStation 12B via Python scripting, benchmarked against the Monte Carlo engine FRED, and validated against an in vivo methodology using patient-specific respiratory traces and machine log files (Cartechini et al., 22 Sep 2025). The model addresses a limitation of conventional 4D plan evaluation, namely that anatomical variation between 4DCT phases can be assessed without explicitly representing the temporal assignment of spots to respiratory phases, which is required for explicit interplay analysis (Cartechini et al., 22 Sep 2025).

1. Definition and clinical rationale

In pencil beam scanning and IMPT, the beam is delivered spot-by-spot and energy-layer-by-energy-layer. For thoracic targets, respiratory motion can range from a few millimeters up to more than $2$ cm, and the delivery time structure can be comparable to or longer than the breathing cycle. The interplay effect is therefore the temporal mismatch between moving anatomy and the moving beam, with the potential to produce cold spots and hot spots within a single fraction, especially for hypofractionated or single-fraction treatments, small spot sizes, or fast energy switching (Cartechini et al., 22 Sep 2025).

Within this framework, “synthetic breathing” does not denote direct measurement of a patient’s breathing before treatment, but the construction of a plausible library of modeled respiratory traces used prospectively. The stated aim is not to reproduce a specific patient’s future breathing exactly, but to cover a plausible envelope of breathing behaviors. This design is explicitly intended for pre-treatment assessment, when actual treatment-room breathing traces are not yet available (Cartechini et al., 22 Sep 2025).

The motivation is strongest where fractionation may not sufficiently average interplay. The framework is described as particularly relevant for hypofractionation and for large motion greater than $2$ cm. A plausible implication is that, in these regimes, explicit temporal modeling becomes a planning-stage decision tool rather than a retrospective verification aid (Cartechini et al., 22 Sep 2025).

2. Respiratory trace library and temporal mapping

The respiratory library consists of sinusoidal synthetic traces with periods of $2$ s, $3.5$ s, and $5$ s, representing fast, typical, and slow breathing. Initial phase is matched to each of the $8$ 4DCT phases, producing $24$ synthetic breathing curves through the Cartesian product of 3 periods and 8 starting phases (Cartechini et al., 22 Sep 2025).

The planning 4DCT is sorted into 8 phases: $0\%$, $25\%$, $50\%$, $2$0, $2$1 inhale, and $2$2, $2$3, $2$4 exhale. These phases define discrete anatomical states over the breathing cycle. Each synthetic trace provides a continuous respiratory signal $2$5, from which respiratory phase at each time point is inferred (Cartechini et al., 22 Sep 2025).

Temporal mapping is then constructed from machine delivery data. Each control point or spot has a delivery time $2$6 obtained from the log file, and for each spot $2$7 the synthetic respiration trace $2$8 assigns a 4DCT phase $2$9. This explicitly reproduces spot sequence within each energy layer, energy-layer sequence, and switching times between energies as determined by the Mevion S250i beam model (Cartechini et al., 22 Sep 2025).

This step is the formal distinction between the synthetic-breathing-based predictive model and conventional 4D evaluation. Conventional 4D plan evaluation at Maastricht, as cited in the source material, accounts for anatomical variation between 4DCT phases but does not include time and therefore does not explicitly map each delivered spot to a respiratory phase. The predictive model adds time and breathing dynamics to the 4D framework (Cartechini et al., 22 Sep 2025).

3. Dose engine implementation and uncertainty integration

The predictive model was implemented in RayStation 12B with Monte Carlo v5.4 using Python scripting. Its required inputs are the 4DCT with 8 phases, the nominal clinical IMPT plan with control points and spot weights, the 24 synthetic breathing traces, and machine log files obtained for the predictive model via a dry run on a water phantom (Cartechini et al., 22 Sep 2025).

For each breathing scenario, RayStation Monte Carlo recalculates dose separately for each of the 8 respiratory phases using only the spots assigned to that phase, yielding phase-resolved dose distributions $2$0. These are then mapped to a reference phase, specified as 50% exhale, using deformable image registration. The accumulated dose is defined as

$2$1

where $2$2 is the deformation operator from phase $2$3 to the reference phase (Cartechini et al., 22 Sep 2025).

The model incorporates setup and range uncertainty directly into the interplay simulation. Setup uncertainty is modeled as systemic isocenter shifts of $2$4 mm in each Cartesian axis, and range uncertainty as systemic HU–RSP curve perturbation of $2$5. To reduce compute time, the implementation uses a probabilistic worst-case sampling strategy in which, for each breathing scenario, one random combination of $2$6 mm shifts and $2$7 range error is sampled and applied to the phase-specific recalculation (Cartechini et al., 22 Sep 2025).

Dose evaluation is summarized through DVH metrics including $2$8, $2$9, and $3.5$0 for targets such as CTVp and CTVn. For treatment-course simulation, one of the 24 synthetic breathing scenarios is randomly assigned to each fraction, fraction-specific accumulated doses are summed, and the full process is repeated 10 times to estimate mean and standard deviation of DVH metrics over a synthetic course (Cartechini et al., 22 Sep 2025).

4. In vivo validation and benchmark methodology

The in vivo model is structurally identical to the predictive model, but replaces synthetic inputs with actual treatment data. Respiratory traces are measured per fraction using the C-RAD surface imaging system, machine log files are recorded after each fraction on the Mevion S250i system, and the same planning 4DCT serves as the anatomical model. The fraction-specific breathing trace is aligned to the 4DCT phase at treatment start, with the explicit caveat that this phase alignment is an assumption (Cartechini et al., 22 Sep 2025).

Quantitative comparison between predictive and in vivo distributions uses the Relative Difference Error (RDE), following Pastor-Serrano et al. For a DVH quantity of interest $3.5$1, the distribution is represented by the percentile vector

$3.5$2

and the comparison is

$3.5$3

with $3.5$4 in this application (Cartechini et al., 22 Sep 2025).

An independent benchmark was performed against the GPU-based Monte Carlo engine FRED v3.76.4 using a clinically validated Mevion S250i beam model. FRED loads all 8 CT phases simultaneously and dynamically activates the appropriate 4DCT phase for each spot according to the time-phase mapping, scoring motion-resolved dose. FRED phase doses are then imported into RayStation for deformable accumulation and DVH analysis (Cartechini et al., 22 Sep 2025).

The reported quantitative results are narrow. Agreement between TPS and FRED showed less than $3.5$5 mean dose difference, and for the lung in vivo case the median absolute difference between FRED and RayStation remained below $3.5$6 for all DVH metrics, with only one fraction showing a deviation greater than $3.5$7 for $3.5$8. Predictive versus in vivo distributions gave RDE less than $3.5$9 for all $5$0, $5$1, and $5$2 metrics in both lung and esophageal cases, while cumulative predictive and in vivo DVH curves agreed within approximately $5$3 for CTVp (Cartechini et al., 22 Sep 2025).

5. Demonstrated clinical cases and fractionation behavior

The methodology was demonstrated on two clinical cases treated on a Mevion S250i system in free breathing without rescanning: one lung cancer case and one esophageal cancer case (Cartechini et al., 22 Sep 2025).

In the lung case, GTV amplitude, defined as maximum center-of-mass displacement across 4DCT phases, was $5$4 mm. The plan used 3 posterior beams at gantry angles of approximately $5$5, $5$6, and $5$7, with robust 3D optimization on an ITV defined as the union of CTV across all phases plus a 1 mm margin. CTVp corresponded to the primary tumor, while CTVn denoted the nodal region, optimized without ITV expansion or specific motion management (Cartechini et al., 22 Sep 2025).

Fractional DVHs in that lung case showed moderate interplay variations per fraction but remained within clinically acceptable limits. For CTVp, cumulative $5$8, $5$9, and $8$0 converged within $8$1 of the nominal plan after approximately 5 fractions, indicating that interplay was effectively averaged out early in the course. For CTVn, a slight persistent underdosage remained at end of course, with $8$2 about $8$3 below nominal and $8$4 about $8$5 below nominal (Cartechini et al., 22 Sep 2025).

In the esophageal case, the plan also used 3 posterior beams of the same geometry, but ITV expansion was 3 mm rather than 1 mm, together with robust optimization for $8$6 range and $8$7 mm setup. This case included a single CTVp. Both predictive and in vivo cumulative DVHs for CTVp stayed within $8$8 of the planned dose from the first fraction, and the plan was described as highly robust under free breathing with large spot sizes even without rescanning (Cartechini et al., 22 Sep 2025).

These findings support a specific clinical interpretation. For primary CTVs optimized on ITV, interplay effects combined with range and setup uncertainty caused only minor dose degradation in the demonstrated cases. Nodal CTV without motion margins was more vulnerable. This suggests that ITV-based robust planning can mitigate interplay effectively for primary thoracic targets under standard fractionation, whereas nodal targets without explicit motion margins warrant closer DVH-based scrutiny (Cartechini et al., 22 Sep 2025).

6. Relation to interplay-robust planning and model scope

The predictive framework is an evaluation and validation methodology rather than a robust optimizer. It predicts dose consequences of delivery time structure, respiratory motion, and sampled setup/range errors for a given plan, and then supports in vivo confirmation once treatment begins (Cartechini et al., 22 Sep 2025). A related but distinct line of work is interplay-robust optimization, where motion scenarios are embedded directly in the optimization problem (Bengtsson et al., 2024).

In the related IPRO framework, synthetic 4DCTs are created by deforming a reference CT using motion patterns obtained from 4DMRI, and motion scenarios are generated by randomly concatenating breathing cycles with variable period and amplitude. In that study, IPRO and IPRO-1C improved near-worst-case target coverage compared to 4DRO, and after normalization to equal target coverage, IPRO with 49 scenarios reduced near-worst-case OAR dose by an average of $8$9, while IPRO-1C with 9 scenarios reduced OAR dose by $24$0 (Bengtsson et al., 2024).

The connection is methodological rather than terminological. Both approaches make respiratory uncertainty explicit through synthetic breathing or synthetic motion scenarios, and both target the temporal interplay between scanned proton delivery and moving anatomy. The present predictive model, however, is anchored to RayStation scripting, dry-run log files, and in vivo validation against actual treatment-room respiratory traces and fraction logs, whereas IPRO formulates a min-max optimization over motion scenarios (Cartechini et al., 22 Sep 2025).

A plausible implication is that the predictive model can serve as a planning-stage gatekeeper even in centers not yet using interplay-robust optimization. It enables prospective risk assessment for questions such as whether rescanning is needed, whether hypofractionation is acceptable for a given motion amplitude, or whether additional margins or different beam angles are required (Cartechini et al., 22 Sep 2025).

7. Strengths, limitations, and broader terminological context

The principal strengths reported for the framework are its pre-treatment predictive capability, its use of a 24-trace synthetic breathing library, full scripting and clinical integration in RayStation, explicit temporal modeling of spot delivery, and unified incorporation of respiratory motion, $24$1 range uncertainty, and $24$2 mm setup uncertainty. Validation against both in vivo measurements and an independent Monte Carlo engine supports internal consistency and clinical feasibility (Cartechini et al., 22 Sep 2025).

Its main limitations are equally explicit. Interplay is evaluated on the planning 4DCT for all fractions, so anatomical evolution such as tumor shrinkage, weight loss, or changes in motion amplitude is ignored. Synthetic traces are idealized sinusoidal traces with constant period inside each scenario, and the in vivo framework assumes alignment between treatment-start respiratory trace and a known 4DCT phase. The delivery time model is also system-specific, depending on the Mevion S250i characteristics, including fixed 230 MeV, 18 range shifters, and relatively large spot sizes (Cartechini et al., 22 Sep 2025).

The phrase “synthetic-breathing-based predictive model” also appears in other arXiv contexts with different meanings. In spontaneous breathing trial prediction, it has been used conceptually for machine-learning systems that map respiratory-flow and ECG-derived features to weaning outcomes (Gonzalez et al., 3 Mar 2025). In remote pulmonary monitoring, related work has used RGB or thermal video plus metadata as synthetic proxies for breathing to predict spirometric indices (Sharshar et al., 29 Jan 2025). These usages share the idea of replacing direct physiological measurement with modeled or proxy respiratory representations, but they are not the same methodology as the thoracic IMPT interplay framework (Cartechini et al., 22 Sep 2025).

In its specific radiotherapy meaning, the term denotes a rigorously defined 4D dose engine that uses synthetic sinusoidal respiratory traces, actual delivery timing from log files, phase-resolved Monte Carlo dose calculation, deformable accumulation, and DVH-based robustness metrics to predict and then validate interplay and motion robustness in free-breathing IMPT. A larger cohort study was stated to be ongoing (Cartechini et al., 22 Sep 2025).