Robust planning against the Achilles heel of proton therapy
Protons deliver almost all of their dose in a narrow Bragg peak, then stop abruptly. That sharp distal falloff is proton therapy's greatest strength — sparing tissue beyond the tumor — and also its greatest vulnerability: any error in predicting exactly where the beam stops shifts the entire high-dose region, not just its edge.
Photon beams deposit dose that decays gradually and monotonically with depth; a small error in tissue attenuation shifts the dose distribution only slightly and continuously — the beam simply keeps going, just with a marginally different intensity at depth. Protons behave completely differently: they travel at a well-defined, finite range and then stop within a few millimeters.
Because the entire therapeutic dose sits in that terminal Bragg peak (or spread-out Bragg peak, SOBP, formed by summing many pristine peaks of different energies), any systematic error in predicting where the proton stops moves the whole high-dose region forward or backward along the beam path. A 3–5mm range error is not a minor perturbation — it can mean the difference between full tumor coverage and a geometric miss, or between sparing and irradiating an adjacent critical structure.
For photons, range error is a dosimetric nuisance. For protons, range error is a geometric miss: the sharp distal edge means uncertainty in physics directly becomes uncertainty in anatomy.
Treatment planning computes proton range from a planning CT scan, but CT scanners measure X-ray attenuation (Hounsfield Units, HU), not proton stopping power. A calibration curve — the "HU-to-relative-stopping-power" (HU→RSP) lookup — converts each voxel's HU value into the stopping-power ratio (SPR) needed to compute range.
This conversion is inherently approximate: two tissues with identical X-ray attenuation can have different elemental composition and therefore different proton stopping power (a phenomenon geometrically distinct from photon physics). Bone, lung, and metal implants are particularly problematic. This single calibration step is generally considered the dominant contributor to range uncertainty, historically responsible for roughly half of the total budget.
Dual-energy CT (DECT) and proton CT (pCT) are emerging technologies that reduce this uncertainty at the source: DECT better separates density from atomic number using two X-ray spectra, while pCT measures proton stopping power directly by imaging with a low-dose proton beam, potentially cutting range uncertainty margins roughly in half.
Even with a perfect calibration, range uncertainty grows from day-to-day variability between the planning CT and the patient on the treatment couch:
• Patient setup error: sub-millimeter to few-millimeter positioning variation despite image guidance, shifting the effective path length through tissue • Weight loss / tumor shrinkage: over a multi-week course, body contour and target volume change, altering the water-equivalent path length • Bowel gas and filling variation: air pockets have drastically different stopping power than soft tissue, and their position is highly variable day to day — a major concern for pelvic and abdominal sites • Respiratory motion: for lung and liver targets, breathing changes both anatomy and the path length the beam traverses
These anatomical sources compound the calibration uncertainty and are why proton centers increasingly use adaptive replanning — re-imaging and re-optimizing the plan when anatomy drifts beyond a clinical threshold.
The most intuitive way to plan a proton beam is to trust the nominal CT and place the Bragg peak (or SOBP) exactly at the tumor edge — the geometrically "optimal" solution if range were known perfectly. In practice this naive, single-scenario approach is fragile: the moment reality deviates from the planning CT, the plan can fail catastrophically rather than degrade gracefully.
In conventional photon radiotherapy, uncertainty is handled by geometrically expanding the clinical target volume (CTV) into a planning target volume (PTV) — a margin that accounts for setup error and organ motion, on the assumption that the dose distribution itself is stable and simply needs to "cover" a slightly larger geometric volume.
For protons, that assumption breaks down. A geometric PTV margin says nothing about range error along the beam direction — it does not know that a 3.5% range overshoot pushes the Bragg peak's sharp distal edge deeper into normal tissue, or that an undershoot pulls it short of the tumor entirely. Early proton planning applied PTV-style geometric margins directly, which is now understood to be an inadequate and sometimes misleading approximation for pencil-beam scanning proton therapy.
A PTV margin blurs a static dose cloud outward. A range shift translates the entire high-dose region. These are fundamentally different types of uncertainty, and only one of them is captured by geometric margins.
Consider a single-scenario plan whose SOBP distal edge is placed to conform exactly to the tumor's distal boundary, with the proximal edge similarly tight. If the true range is shorter than planned (range undershoot, e.g. due to higher-than-modeled stopping power), the entire dose distribution falls short: a cold spot opens up at the tumor's distal margin, exactly the region most likely to harbor microscopic disease.
Conversely, if the true range is longer than planned (range overshoot), the Bragg peak — and its characteristically sharp, high-gradient falloff — lands beyond the tumor, delivering a full-dose "hot spot" to whatever normal tissue or organ-at-risk sits just past the target. Because the falloff is only a few millimeters wide, this is an abrupt, binary-feeling failure rather than a gradual dose gradient.
The clinical stakes of this fragility are high in both directions. Underdosing the distal tumor margin risks marginal-miss local recurrence, particularly dangerous because it is the deep/distal margin — often adjacent to the most critical structure the beam is trying to avoid — that is most exposed.
Overdosing normal tissue just beyond the target risks toxicity in exactly the organ the proton beam was chosen to spare: spinal cord myelopathy, brainstem injury, rectal or bladder toxicity, or pneumonitis, depending on site. This dual-sided risk — recurrence on one side, toxicity on the other — is why single-scenario, margin-only proton planning has been superseded by robust optimization at most modern proton centers for sites with meaningful range uncertainty.
Rather than trusting a single nominal dose calculation, robust planning explicitly enumerates the plausible ways reality can deviate from the plan — shifted ranges, shifted setup positions, and their combinations — and computes the dose distribution for every one of them. The result is not one curve but an envelope of possible outcomes.
The scenario-based robust framework, formalized in proton therapy by Pflugfelder, Unkelbach and colleagues in the early 2010s, replaces the single nominal dose calculation with a discrete set of "error scenarios." Each scenario represents a self-consistent, physically plausible combination of systematic range error and patient setup shift.
A typical implementation perturbs the range by the clinical uncertainty budget (commonly ±3.5% of the nominal range, reflecting calibration uncertainty) and perturbs the isocenter position by the setup uncertainty (commonly 1–5mm) along each principal axis. Combined combinatorially — range shifted high/low/nominal crossed with setup shifted in multiple directions — this typically yields 9 to 21 scenarios per beam, all of which are optimized and evaluated together rather than sequentially.
Each scenario requires an independent proton dose calculation: the stopping-power map is shifted according to the assumed range error, the beam isocenter is shifted according to the assumed setup error, and a full pencil-beam or Monte Carlo dose calculation is performed on that perturbed geometry.
This is computationally expensive — a 9-scenario robust optimization costs roughly nine times the dose calculations of a conventional single-scenario plan, and modern robust workflows with 15-21 scenarios scale accordingly. Advances in GPU-accelerated Monte Carlo dose engines have been essential in making scenario-based robust optimization practical for routine clinical use rather than a research curiosity.
A key conceptual insight from Unkelbach's work is that the "worst case" should be evaluated voxel by voxel, not scenario by scenario. A single global worst-case scenario does not exist — the scenario that produces the lowest dose to a proximal tumor voxel is generally not the same scenario that produces the lowest dose to a distal tumor voxel.
Voxel-wise worst-case dose is therefore constructed by, for each voxel independently, taking the minimum (for target coverage objectives) or maximum (for OAR sparing objectives) dose across all scenarios. This produces the true envelope of worst-case outcomes shown as the shaded band on the depth-dose curve, and is the objective that the optimizer in Stage 4 directly minimizes against.
This is the technical crux of range-robust planning: optimize against the worst dose seen by each individual voxel across all scenarios, not against any single "worst" scenario as a whole.
Modern intensity-modulated proton therapy (IMPT) delivers dose as thousands of individually weighted pencil-beam spots. Robust optimization exploits this flexibility: instead of tuning spot weights to satisfy one nominal dose distribution, the optimizer tunes them so that every scenario in the uncertainty set simultaneously satisfies clinical objectives.
Robust optimization reformulates the planning problem itself. Instead of "expand the target geometrically, then optimize dose to the expanded volume" (the PTV approach), the objective function directly incorporates every uncertainty scenario: for each iteration, the optimizer evaluates how well the current spot weights perform not just on the nominal geometry but on all ~9-21 perturbed geometries simultaneously, and adjusts weights to improve the worst-performing scenario for each clinical objective (a minimax or worst-case formulation).
The practical effect is that the optimizer no longer needs an artificial geometric margin at all — the CTV itself, not an expanded PTV, is the optimization target, because robustness against range and setup error is now built directly into the objective function rather than approximated by geometry.
Concretely, each pencil-beam spot is defined by an energy (which determines its individual Bragg peak depth) and a monitor-unit weight (which determines its intensity). Thousands of overlapping spots at different energies are summed to build the composite SOBP.
During robust optimization, spots near the tumor's proximal and distal edges receive extra scrutiny: their weights are adjusted — some increased, some decreased — so that the summed dose from all spots remains acceptable across every scenario's shifted geometry. In practice this often means slightly extending the effective coverage of edge spots and softening the sharpest nominal conformality, trading a small amount of nominal-scenario "tightness" for much greater resilience when reality deviates from the plan.
A robust plan can look less tightly conformal on the nominal CT than a naive plan — and that is by design. The naive plan's tightness is an illusion of precision that a real 3.5%-plus-setup shift immediately breaks.
Robust optimization is not free. Optimizing against 9-21 scenarios simultaneously multiplies the dose-calculation and optimization workload roughly proportionally, requiring the GPU-accelerated dose engines and optimization solvers that have become standard in modern treatment planning systems over the past decade.
There is also a dosimetric trade-off: because the plan must satisfy every scenario, not just the nominal one, the nominal-scenario dose distribution is typically slightly less sharp than an equivalent naive plan — a small, deliberate sacrifice of theoretical best-case conformality in exchange for guaranteed worst-case adequacy. Clinically, this trade is considered strongly favorable: the naive plan's "best case" is never actually delivered once real-world uncertainty is present, while the robust plan's worst case is what patients actually receive on an uncertain day.
A robust plan is only as good as its evaluation. Instead of a single dose-volume histogram (DVH), robust plans are judged on worst-case DVH bands — the envelope of possible outcomes across the full scenario set — for both the tumor and every nearby organ at risk, confirming that coverage and sparing hold up under the full range of plausible range and setup error.
Rather than reporting one DVH curve, robust plan evaluation reports a band: for every dose level, the minimum and maximum volume receiving at least that dose across all scenarios. For the tumor, clinicians check that even the worst-case (lowest) curve in the band still meets coverage goals (e.g., 95-98% of the volume receiving at least 95% of prescription dose). For organs at risk, they check that even the worst-case (highest) curve in the band still respects dose-tolerance constraints.
This voxel-wise worst-case dose — built exactly as described in Stage 3, taking the min or max across scenarios independently at each voxel — is what distinguishes a genuinely robust plan from one that merely looks good on the nominal CT. Side by side, the naive Stage-2 plan shows a band with a visible gap or breach at the tumor edge; the robust plan's band stays within bounds throughout.
The single most important sanity check in proton QA: does the worst-case curve of the band still meet the clinical goal? If only the nominal curve meets it, the plan is not actually safe to deliver.
Robust optimization protects against the uncertainty budget it was designed for, but it cannot protect against anatomical changes that exceed that budget — a patient who loses significant weight, a tumor that shrinks substantially, or a persistent shift in bowel gas pattern can push the true anatomy outside the modeled scenario envelope.
Modern proton workflows address this with adaptive replanning: periodic verification imaging (repeat CT, cone-beam CT, or in-room CT-on-rails) is compared against the planning CT, and if anatomical or dosimetric drift exceeds a pre-defined clinical threshold, the plan is recalculated or re-optimized on updated anatomy. Sites with high anatomical variability — head and neck (weight loss, tumor response), lung (breathing motion, atelectasis changes), and pelvis (bladder/rectal filling, bowel gas) — are the most frequent candidates for adaptive intervention.
Robust optimization manages range uncertainty after the fact; a complementary research and clinical direction is reducing the uncertainty itself before planning even begins:
• Dual-energy CT (DECT): using two X-ray spectra to separately estimate electron density and effective atomic number improves the HU-to-stopping-power conversion, potentially reducing the calibration contribution to the range-uncertainty budget • Proton CT (pCT) and proton radiography: imaging with a low-dose diagnostic proton beam measures stopping power directly along the proton path, bypassing the X-ray-to-proton conversion step entirely • In-vivo range verification: positron emission tomography (PET) imaging of beam-induced positron-emitter activation, and prompt-gamma imaging of instantaneous nuclear emissions during treatment, both offer ways to confirm the delivered range matches the planned range
As these technologies mature and enter routine clinical use, the standard 3.5%-plus-setup margin recipe is expected to shrink — but until uncertainty is eliminated at the source, scenario-based robust optimization remains the clinical standard for guaranteeing that a proton plan performs safely, not just optimally.