Quantification of plan robustness against different uncertainty sources for classical and anatomical robust optimized treatment plans in head and neck cancer proton therapy

Abstract

Objective:

Classical robust optimization (cRO) in intensity-modulated proton therapy (IMPT) considers isocenter position and particle range uncertainties; anatomical robust optimization (aRO) aims to consider additional non-rigid positioning variations. This work compares the influence of different uncertainty sources on the robustness of cRO and aRO IMPT plans for head and neck squamous cell carcinoma (HNSCC).

Methods:

Two IMPT plans were optimized for 20 HNSCC patients who received weekly control CTs (cCT): cRO, using solely the planning CT, and aRO, including 2 additional cCTs. The robustness of the plans in terms of clinical target volume (CTV) coverage and organ at risk (OAR) sparing was analyzed considering stepwise the influence of (1) non-rigid anatomical variations given by the weekly cCT, (2) with fraction-wise added rigid random setup errors and (3) additional systematic proton range uncertainties.

Results:

cRO plans presented significantly higher nominal CTV coverage but are outperformed by aRO plans when considering non-rigid anatomical variations only, as cRO and aRO plans presented a median target coverage (D_98%) decrease for the low-risk/high-risk CTV of 1.8/1.1 percentage points (pp) and −0.2 pp/−0.3 pp, respectively. Setup and range uncertainties had larger influence on cRO CTV coverage, but led to similar OAR dose changes in both plans. Considering all error sources, 10/2 cRO/aRO patients missed the CTV coverage and a limited number exceeded some OAR constraints in both plans.

Conclusion:

Non-rigid anatomical variations are mainly responsible for critical target coverage loss of cRO plans, whereas the aRO approach was robust against such variations. Both plans provide similar robustness of OAR parameters.

Advances in knowledge:

The influence of different uncertainty sources was quantified for robust IMPT HNSCC plans.

Introduction

Intensity-modulated proton therapy (IMPT) has shown to be promising for the treatment of head and neck squamous cell carcinoma (HNSCC), due to its high-dose conformity and reduced dose to the normal tissue in comparison with photon-based intensity-modulated radiation therapy (IMRT).^1–4 However, protons are more sensitive to uncertainties, e.g. patient setup and particle range, and it has been shown that a simple planning target volume (PTV) expansion from the clinical target volume (CTV) is not sufficient to account for those uncertainties.^5–7

Robust IMPT optimization algorithms, which include uncertainties in proton range and rigid isocenter shifts to mimic patient setup variability, have been widely investigated.^8–11 In general, these robust plans do not consider potential non-rigid anatomical variations, which may occur during the course of treatment, such as random variations in patient posture and filling of cavities, or systematic tumor shrinkage and weight changes. The latter ones are hard to predict prior treatment while different postures might be imaged beforehand and cavity densities could be overwritten in the planning CT. As the daily patient geometry is a result of the superposition of all variations and as their individual dosimetric influences cannot clearly be differentiated, it might be worthwhile to include any additional information about non-rigid anatomical variability available before treatment into the optimization for increasing the plan robustness as shown by novel studies.^12–14

The evaluation of the plan robustness, i.e. the potential dose distortion due to uncertainties, should be assessed as part of the overall plan evaluation procedure.¹⁵ In general, robustness analysis methods calculate perturbed doses with defined setup and range uncertainty values.^5,16–19 However, anatomical variations during the treatment course are usually not considered as an additional variable. Moreover, the individual contribution of the different sources of uncertainty in the overall plan robustness is in general not assessed.

In this retrospective work, we compare the influence of different uncertainty sources on the robustness of IMPT plans for HNSCC generated by two different planning approaches, being the classical robust optimization (cRO), that takes rigid setup and range uncertainties into account, and by anatomical robust optimization (aRO), which additionally considers non-rigid patient positioning variations, proposed in a previous study.¹⁴ In the current work, a comprehensive robustness analysis method was applied stepwise by considering the impact on the overall plan robustness of (1) non-rigid anatomical variations during the treatment course, (2) fractionwise rigid random setup errors and (3) systematic range errors.

Methods and materials

Patient data and treatment planning

Data sets from 20 locally advanced HNSCC patients were included in our in silico study, 17 of whom were treated with IMRT, 2 with double scattered (DS) proton radiotherapy and 1 with a mixed (IMRT/DS) treatment at the University Hospital Carl Gustav Carus, Dresden, Germany, between August 2015 and July 2016. These were the patients with data sets consisting of a planning CT and weekly scheduled control CTs (cCT) acquired with adequate length during the course of treatment (median: 6, range: 4–7). All 17 IMRT patients received the regular cCTs because of their large tumor volume and/or a potential change of the external in case of tumor shrinkage. A high-risk CTV including the primary tumor, surgical tumor bed and potential metastatic lymph nodes and a low-risk CTV electively including the bilateral cervical lymph nodes were contoured by an experienced radiation oncologist. Spinal cord, brainstem, parotid glands, larynx, oral mucosa, pharyngeal constrictor muscles and esophageal inlet muscle were delineated as organs at risk (OARs). After contours were transferred from the planning CT to each registered cCT using a deformable image registration (DIR) algorithm,²⁰ the same radiation oncologist reviewed and corrected them if necessary. Anatomical variations between the planning and control CTs are presented for exemplary patients in section C of the Supplementary Material 1.

Prescribed mean doses of 57 Gy to the low-risk CTV and 70 Gy to the high-risk CTV were planned to be delivered with a simultaneous integrated boost (SIB) technique in 33 fractions. An additional intermediate dose region between low- and high-risk volume of 10 mm was generated to assure a steep SIB dose gradient.^18,21 The plans were optimized to deliver the prescribed dose to the CTVs with the following planning objectives: D_98% ≥ 95% and D_2% ≤ 107% of the prescribed dose, where D_98% and D_2% are the minimum doses to 98% and 2% of the target volume, respectively. Maximum and mean dose (D_max, D_mean) objectives to the OARs were defined as follows: D_max(spinal cord) < 45 Gy, D_max(brainstem) < 54 Gy, D_mean(parotid glands) ≤ 26 Gy, D_mean(larynx) < 40 Gy, D_mean(constrictor muscles) < 42 Gy. Mean doses to the oral mucosa and esophageal inlet muscle were reduced as low as reasonably achievable. Only the OAR volumes outside the CTV were considered during the optimization process.

Two IMPT plans using robust optimization were calculated in RayStation v5.99 (RaySearch Laboratories AB, Stockholm, Sweden): cRO, using the planning CT for the optimization, and aRO, considering for the plan optimization several planning CTs which represent random non-rigid patient positioning variability. As such additional planning CTs were not available, the first two control CTs were included in the optimization process, since they present random non-rigid positioning variations without relevant treatment-induced anatomical changes compared to the last cCT, as shown in Supplementary Material 1.

Both optimizations took 3 mm setup and 3.5% range uncertainty into account, optimizing via the minimax approach, which considered 21 different scenarios in the optimization in the cRO case.^{10,18,21–25} Since two additional CTs were included in the aRO plans, 3 × 21 = 63 scenarios were considered in this case. The plans were generated using three beams with the same configuration (gantry/couch angles of 180°/0°, 60°/340° and 300°/20°), a minimum air gap of 3 cm and a range shifter of 7.5 cm water equivalent thickness, and considering a constant relative biological effectiveness (RBE) for proton beams of 1.1. Objective functions for both CTVs, spinal cord, brainstem and parotid glands were selected as robust. The generated nominal dose distributions are referred to as D_Nom.

Robustness analysis considering anatomical variations and setup and range uncertainties

Total cumulative doses (D_Cum) considering anatomical variations were generated by recalculating the plans on each weekly cCT, followed by a non-rigid deformation of the calculated dose to the planning CT for dose accumulation.

A procedure (schematically depicted in Figure 1) has been presented previously²⁶ for evaluating the additional influence of setup and range uncertainties on the plan robustness in combination with anatomical variations. First, 10 perturbed dose distributions with random rigid setup errors were generated by drawing for each treatment fraction n (n = 1, 2, …, 33), the isocenter shift for each cardinal direction (x_n, y_n, z_n) from a Gaussian distribution with mean μ = 0 mm and standard deviation σ = 1.5 mm^17,18,27 and considering in each fraction the respective anatomy of the weekly cCT. The 33 single-fraction perturbed doses with different random setup errors were calculated, deformed to the planning CT and summed to generate 1 of the 10 cumulative perturbed dose distributions D_Per0. Second, range errors of +3.5% and -3.5% were included in the perturbed dose calculation with the same random setup errors per fraction already generated, repeating the same procedure described before for each range error, resulting in a total of 20 additional cumulative perturbed dose distributions D_PerR (10 with range error of +3.5% and 10 with range error of -3.5%).

Figure 1.

Workflow for robustness analysis by generating 30 cumulative perturbed dose distributions via 10 sets of 33 fractionwise random setup errors and three systematic range errors.

We computed for each group of cumulative perturbed doses the worst-case values, which are the minimum value for the CTV D_98% and the maximum value for the CTV D_2%, the OAR dose volume histogram parameters and the integral dose in the normal tissue.²⁸ To evaluate the influence of the different factors on the plan robustness, worst case and D_Cum dose volume histogram parameters were compared to those from the nominal dose (D_Nom) for anatomical variations alone (D_Cum), anatomical variations plus random setup uncertainty (D_Per0) and finally for anatomical variations plus setup and range uncertainties (D_PerR). The number of patients exceeding any OAR or a CTV coverage constraint due to the influence of the different error sources were extracted. Moreover, the perturbation widths for the CTV and OAR dose parameters were calculated from the 30 cumulative perturbed scenarios, defined as the difference between the maximum and minimum value, and compared for both planning approaches.

Statistical analysis

To evaluate differences between the two plan approaches and between the uncertainty sources over the patient cohort, Wilcoxon signed-rank tests were performed in SPSS v25 (IBM Corporation, New York, SA), considering a p-value < 0.05 as statistically significant. 

Results

Nominal and total cumulative doses

Both nominal plan doses D_Nom fulfilled the target coverage of D_98% ≥ 95%, with significantly (p < 0.001) higher values for the cRO plans. The median (minimum) values for the low- and high-risk CTVs were 98.2% (96.6%) and 97.9% (96.8%), compared to 97.5% (95.2%) and 97.4% (95.4%) for the aRO approach, respectively (Table 1, Figure 2). Nominal OAR dose parameters were similar for both plans except the slight but significantly higher mean doses to the oral mucosa (p = 0.006) and esophageal inlet muscle (p = 0.028) in the aRO plans (Table 1, Figure 3). The dosimetric constraints could not be met in a number of cases for the ipsilateral parotid gland (8× cRO, 8× aRO), the larynx (9× cRO, 8× aRO) and pharyngeal constrictor muscles (15× cRO, 15× aRO) due to their vicinity to the large target structures. Integral doses to the normal tissue were 8% higher for the aRO plans (p < 0.001), with a median of 112.84 Gy∙l, compared to 104.72 Gy∙l for the cRO approach.

Table 1.

CTV (upper part) and OAR (lower part) dose parameters of the nominal plans and their degradation due to the stepwise influence of anatomy variation (D_Nom - D_Cum), setup errors (D_Cum - D_Per0) and range errors (D_Per0 - D_PerR), and the cumulative influence of all three aspects (D_Nom-  D_PerR). Stated numbers are the median (range) values for the whole patient cohort. A positive value indicates a reduction of the respective dose parameter by the particular error source. Values of the cRO or the aRO plans are marked by an asterisk (*) if they present significantly better values in the nominal dose distribution (D_Nom) or lower degradation by the respective error source(s) compared to the other planning approach (Wilcoxon signed-rank test, level of significance: 0.05).

ROI	Metric	Plan	DNom	DNom - DCum	DCum - DPer0	DPer0 - DPerR	DNom - DPerR
CTV low–risk	D98% (%)	cRO aRO	98.2 (96.6–99.3)*97.5 (95.2–98.8)	1.8 (0.1–9.2)-0.2 (-0.7–4.2)*	0.5 (0.0–1.9)0.2 (-0.2–1.0)*	0.6 (0.1–2.8)0.4 (-0.1–0.9)*	2.9 (0.6–11.5)0.6 (-0.4–6.0)*
	D2% (%)	cRO aRO	107.1 (104.7–110.0)*107.6 (103.6–110.8)	0.2 (-1.2–1.1)0.7 (-0.7–1.5)*	0.2 (-0.2–0.6)0.3 (-0.2–0.5)	-0.1 (-0.6–0.1)-0.2 (-0.5–0.1)	0.2 (-1.4–1.2)0.7 (-1.2–2.0)*
CTV high–risk	D98% (%)	cRO aRO	97.9 (96.8–98.8)*97.4 (95.4–98.5)	1.1 (-0.3–7.9)-0.3 (-1.9–2.1)*	0.5 (0.0–2.1)0.3 (0.0–2.4)*	0.3 (0.0–1.5)0.2 (-0.1–0.7)*	1.7 (-0.1–9.9)0.1 (-1.6–2.5)*
	D2% (%)	cRO aRO	103.9 (102.0–105.5)103.9 (100.6–105.9)	0.5 (-1.2–1.0)0.6 (-0.9–1.3)	0.2 (-0.1–0.5)0.2 (0.0–0.4)	-0.1 (-0.4–0.1) 0.0 (-0.4–0.2)	0.6 (-1.0–1.1)0.7 (-0.9–1.3)
Spinal cord	D1cc (Gy)	cRO aRO	24.9 (11.8–31.4)23.8 (12.0–33.2)	-0.7 (-2.3–0.3)-0.6 (-2.2–0.8)*	-0.3 (-0.9–0.1)-0.3 (-1.0–0.2)	-0.2 (-1.2–0.0)-0.2 (-0.8–-0.1)	-1.2 (-3.6–0.0)-1.1 (-3.4–0.4)*
Brainstem	D1cc (Gy)	cRO aRO	12.4 (0.4–22.9)11.5 (0.7–23.4)	-0.1 (-3.0–0.9)-0.1 (-2.9–1.0)	-0.4 (-1.2–0.1)-0.4 (-1.1–0.2)	-0.1 (-1.6–0.0)-0.1 (-1.4–0.0)	-1.0 (-4.2–0.2)-0.7 (-4.3–0.4)
Parotid glandipsilateral	Dmean (Gy)	cRO aRO	21.2 (19.2–55.2)21.0 (16.7–54.4)	-1.0 (-3.1–1.6)-0.9 (-3.7–1.4)	-0.8 (-4.3–1.4)-0.9 (-1.2–0.0)	-0.2 (-0.7–0.0)-0.2 (-0.6–0.1)	-1.9 (-6.6–1.5)-1.8 (-3.9–1.2)
Parotid glandcontralateral	Dmean (Gy)	cRO aRO	20.0 (17.1–21.4)20.0 (10.8–21.3)	0.1 (-4.8–1.7)0.1 (-4.0–1.3)	-0.7 (-3.4–-0.1)-0.7 (-1.2–-0.1)	-0.2 (-0.4–0.0)*-0.2 (-0.7–0.0)	-0.7 (-6.1–1.0)-0.7 (-5.5–0.5)
Larynx	Dmean (Gy)	cRO aRO	36.6 (23.7–69.9)35.3 (24.3–69.8)	-0.4 (-12.9–2.7)*-0.9 (-13.7–2.6)	-0.3 (-1.1–0.1)-0.4 (-1.1–0.0)	-0.4 (-1.1–0.0)-0.4 (-1.1–0.0)	-1.1 (-13.8–2.5)*-1.6 (-14.8–1.8)
Oral mucosa	Dmean (Gy)	cRO aRO	38.7 (17.2–65.4)*40.0 (17.5–65.3)	0.1 (-3.8–3.5)0.0 (-3.8–3.2)	-0.3 (-0.7–0.3)*-0.3 (-0.7–0.2)	-0.8 (-1.5–-0.2)-0.7 (-1.2–-0.1)*	-0.8 (-4.2–2.4)-0.9 (-4.3–2.5)
Pharyngealconstrictor muscles	Dmean (Gy)	cRO aRO	50.6 (39.4–64.4)50.9 (40.3–64.4)	-0.4 (-3.4–1.1)*-0.8 (-3.2–0.9)	-0.2 (-0.7–-0.1)-0.2 (-0.8–0.0)	-0.6 (-0.9–-0.1)-0.4 (-1.1–-0.1)	-1.2 (-4.1–0.6)*-1.5 (-4.0–0.4)
Esophageal inlet muscle	Dmean (Gy)	cRO aRO	38.2 (16.2–69.7)*38.5 (21.8–69.3)	-0.5 (-4.2–5.1)-0.9 (-4.8–6.5)	-0.4 (-0.9–0.6)-0.4 (-0.9–0.6)	-1.1 (-2.1–-0.1)-0.9 (-2.0–-0.1)	-2.0 (-5.9–4.0)-2.2 (-6.5–5.1)
Healthy tissue:External-CTV	Integral dose(Gy∙l)	cRO aRO	104.7 (66.4–133.3)*112.8 (68.8–146.1)	0.5 (-9.3–6.9)0.5 (-13.6–7.6)*	-0.1 (-8.4–0.8)-0.1 (-9.6–0.7)	-2.1 (-3.2–-0.9)-1.8 (-3.5–-0.9)	-0.9 (-20.4–4.6)*-1.2 (-25.3–6.0)

cRO, classical robust optimization; aRO, anatomical robust optimization; ROI, region of interest; CTV, clinical target volume; D_98%, dose to the 98% of the volume; D_2%, dose to the 2% of the volume; D_1cc, near maximum dose to the 1 cc of the volume; D_mean, mean dose.

Figure 2.

Target coverage statistics of the whole patient cohort for plans with classical robust optimization (cRO, blue) and with anatomical robust optimization (aRO, yellow) in nominal dose (D_Nom), total cumulative dose (D_Cum) and worst-case of the 10 perturbed cumulative dose distributions without range uncertainty (D_Per0) and of the 20 perturbed cumulative dose distributions considering range uncertainty (D_PerR). The boxes present the interquartile range and the horizontal lines the median values. Whiskers include data within the 1.5-fold of the interquartile range; values beyond are outliers.

Figure 3.

Comparison of organ at risk dose parameters and integral dose to the healthy tissue for plans with classical robust optimization (cRO, blue) and anatomical robust optimization (aRO, yellow) in the nominal dose (D_Nom), total cumulative dose (D_Cum) and worst-case of the 10 perturbed dose distributions without range uncertainty (D_Per0) and of the 20 perturbed dose distributions considering range uncertainty (D_PerR) for all patients. Dashed lines indicate planning objectives, if applicable.

When considering only anatomical variation over the course of treatment, the target coverage dropped significantly in the total cumulative doses D_Cum for the cRO plans: the median (maximum) target coverage loss was 1.8 (9.2) pp (percentage points) and 1.1 (7.9) pp for the low- and high-risk CTV, respectively (Table 1). In five patients (25%), the D_98% value was even below the threshold of 95% for at least one of the CTVs, i.e. a plan adaptation would have been required under real treatment conditions. Conversely, median values of the aRO plans showed no target coverage loss (-0.2 pp and -0.3 pp for low- and high-risk CTV, respectively). However, one patient (Patient 1) showed loss in coverage by 4.2 pp and 1.2 pp to values below the clinical objective (92.4% and 94.2% for each CTV, respectively) due to severe anatomical changes during the treatment course (Figure 2). The median increase of all OAR dose parameters in the D_Cum was ≤ 1 Gy (Table 1) and was significant for the spinal cord (cRO: p < 0.001; aRO: p = 0.005), ipsilateral parotid gland (cRO: p = 0.008; aRO: p = 0.01) and pharyngeal constrictor muscles (aRO: p = 0.008). Comparing between the plans, a larger increase of the cRO near maximum dose of the spinal cord (p = 0.04), and the aRO mean dose of the larynx (p = 0.02) and pharyngeal constrictor muscles (p = 0.001) was observed. A relevant increase in dose of more than 5 Gy was found for the larynx mean dose in two individual cases for both plans. OAR mean dose constraints were violated due to the non-rigid anatomy variations in a few cases for the ipsilateral parotid gland (1× aRO), larynx (1× cRO, 2× aRO) and pharyngeal constrictor muscles (3× cRO, 3× aRO).

Additional influence of setup and range uncertainties

Compared to D_Cum, the worst case reduction in target coverage due to the inclusion of random rigid setup uncertainties (D_Per0) was significant (p ≤ 0.001) for all CTVs and plan approaches, and were significantly larger for the cRO plans with a median reduction of 0.5 pp for the low- and high-risk CTV compared to the 0.2 pp and 0.3 pp, respectively, for the aRO plans (Table 1). Consequently, the D_Per0 target coverages were significantly worse (p ≤ 0.001) for the cRO plans, with median (minimum) values of 95.6% (87.2%) and 96.3% (88.7%) for the low- and high-risk CTV, respectively, compared to the aRO plans with values of 96.6% (91.5%) and 97.5% (93.9%), respectively (Supplementary Table 1).

Additionally considering the range uncertainty, the reduction of the worst-case low-/‌high-risk CTV coverage from the 20 perturbed doses (D_PerR) was also significant (p < 0.001) for both plans, and was again significantly larger for the cRO plans with median values of 0.6 pp/0.3 pp compared to 0.3 pp/‌0.2 pp in the aRO plans, respectively (Table 1). The median (minimum) worst-case coverage values for all considered error sources were 95.2% (86.1%) and 96% (90.6%) for low- and high-risk CTV in the cRO approach and 96.4% (90.6%) and 97.2% (93.8%) for the aRO plan, respectively (Supplementary Table 1). These numbers correspond to 10 and 2 patients missing the coverage objective for at least one of the CTVs by the cRO and aRO approach, respectively. The target coverage in the entire patient cohort for the different error-influenced dose distributions is depicted in Figure 2.

The range of the target coverage values (D_98%) per patient in the 30 cumulative perturbed dose distributions (cp. perturbation width in Supplementary Table 1) was significantly larger (p ≤ 0.002) for cRO plans, with median (maximum) values of 1.5 (5.5) pp and 1.0 (3.9) pp in the low- and high-risk CTV, respectively, compared to 1.0 (2.9) pp and 0.6 (3.3) pp for the aRO plans, respectively. This larger variation for the cRO cases might indicate a lower robustness.

Only for a few patients, e.g. patients 10 and 20, the cRO approach performed equally well as the aRO plan (Figure 4). Moreover, for the cRO approach, the mean (± standard deviation) number (percentage) of the 30 scenarios per patient fulfilling the clinical target coverage objective was 22.3 ± 12.0 (74.2%) and 22.5 ± 12.0 (74.8%) for the low- and high-risk CTV, respectively, compared to 28.2 ± 6.8 (93.8%) and 28.5 ± 6.7 (95.0%), respectively, for the aRO plans. As an extreme case, for Patient 1 both plan approaches were below the objectives in all scenarios, since this patient presented significant anatomical variations during the treatment course.

Figure 4.

Ranges of the target coverage (D_98%) values in the 30 perturbed dose distributions per patient. The central line in each box represents the median value. The dashed line represents the clinical objective (95%), whereas the small circles represent the nominal plan values D_Nom and the crosses the values from the total cumulative dose D_Cum considering anatomical variations only. aRO, anatomical robust optimization; cRO, classical robust optimization; CTV, clinical target volume.

Regarding the OARs doses (Figure 3, Table 1), a significant dose increase with median values up to 1.1 Gy was observed for setup (D_Per0) and range errors (D_PerR) for all organs and both plan approaches (p ≤ 0.002). One additional patient missed the mean dose constraints of the ipsi- and contralateral parotid gland in the cRO plan when adding setup error influences (D_Per0). Comparing the two plans (Table 1), both had an overall equal robustness concerning the OAR dose parameters. Minor but significantly lower robustness was found against rigid setup errors for the aRO D_mean of oral mucosa (p = 0.005) and against range errors for the aRO D_mean of the contralateral parotid gland (p = 0.04) and the cRO D_mean of oral mucosa (p = 0.005). Comparing the total influence of all error sources relative to the nominal plan (D_Nom - D_PerR), the cRO plans showed significantly better robustness for the D_mean of the larynx (p = 0.02) and pharyngeal constrictor muscle (p = 0.01) and a worse robustness for the D_1cc of the spinal cord (p = 0.002). Individual relevant increases in dose of more than 5 Gy in the D_PerR compared to the nominal plan were found for the D_mean of the ipsi- (1× cRO) and contralateral (1× cRO, 1× aRO) parotid gland, the larynx (2× cRO, 2× aRO) and esophagus inlet muscle (1× cRO, 1× aRO).

The integral dose to the normal tissue was in all cases about 8% higher for the aRO plans, being in agreement with which was observed in the nominal doses. The D_PerR values showed, compared to the nominal value, a significantly lower (p = 0.02) increase for the cRO plans with a median value of 0.9 Gy∙l compared to 1.2 Gy∙l for the aRO plans, respectively. These median values of the cohort are dominated by the range error influence while non-rigid anatomy variations and setup errors could have large individual influence. The perturbation ranges of the OAR parameters and the healthy tissue integral dose were similar for the two plan approaches without any significant difference (Supplementary Table 1), indicating similar robustness against rigid setup and range errors.

Discussion

Including information about potential anatomical variations, such as non-rigid positioning uncertainties, into the robust optimization of IMPT plans can improve the overall plan robustness as it was shown in recent studies.^12–14 In the present study, the influence of different uncertainty sources on the robustness of HNSCC proton plans was for the first time evaluated based on a multiscenario stepwise evaluation. Non-rigid anatomical variations during the whole treatment course played the most critical role in target coverage degradation for classic robust optimized plans. Their influence was significantly reduced by anatomical robust optimization, although the optimization method includes only the more random changes like patient posture and not the systematic changes like tumor shrinkage. Additional rigid setup and range uncertainties had also a significantly lower influence on aRO CTV coverage, but the coverage loss was in the same order of magnitude for both plans, possibly related to the fact that both plans were optimized with the same setup and range error parameters. In total, as shown in Figure 4, the range of CTV D_98% values was in general larger for the cRO cases, i.e. there is an increased variability of the dose parameter when all the uncertainties were considered.

Under usual clinical conditions, only the influence of anatomical changes throughout the treatment is assessed by means of regular control CTs and dose tracking techniques. In such a simplified but fast analysis (comparison between D_Cum and D_Nom), aRO plans would have led to a five-fold lower number of patients requiring plan adaptations compared to the cRO plans (one vs five patients). A clinically applied dose tracking throughout the treatment course that takes random setup errors into account is complex since the fractionation effect lowers the influence of setup errors with increasing number of delivered fractions.^17,18 However, with our determination of the cumulative (33 fractions) perturbed doses considering non-rigid positioning variations plus fractionwise random rigid setup errors and systematic range errors, we ended up with worst-case target coverage values below the clinical objective (D_98% > 95%) in 2 (10%) of the aRO plans and 10 (50%) of the cRO plans (cp. Figure 4), i.e. doubling the numbers compared to the D_Cum analysis and confirming the 1:5 ratio of potential patients requiring plan adaptations between the aRO and cRO optimization strategies. Whether these threshold-based numbers derived from the physical dose distributions would be linked to significantly lower tumor control, when treating without any plan adaptation, remains questionable here.

The price for the improved robustness of aRO plans is a sligthly increased integral dose to the normal tissue by about 8 Gy∙l, but this increase was not necessarily reflected in the OAR dose parameters, which were mostly similar between both plan approaches, showing only for the oral mucosa and esophageal inlet muscle a slight but significantly higher mean dose. Thus, the price of improved plan robustness can be considered as low and acceptable. Due to the similarity of OAR dose parameters and their similar robustness in both plans, one might not expect a relevant increase in normal tissue complications when applying the aRO planning approach under clinical conditions. A potentially improved tumor control compared to a cRO-based non-adapted HNSCC IMPT treatment could be anticipated based on the finding that five-times more patients fulfilled the target coverage constraints by the aRO plans even under the influence of all three error sources. Whether larger range and setup robustness parameters could potentially improve anatomical robustness of the cRO plans without exceeding the OAR constraints and enhancing too much the normal tissue dose was not investigated, and thus remains unproven here.

The aRO approach may be implemented into the clinical workflow by the acquisition of several planning CTs, taken after complete patient repositioning between the scans to acquired realistic non-rigid differences in patient positioning, e.g. variations of head tilt and rotation, shoulder position, neck flexion, mandible and palate position in HNSCC patients. So far, such additional planning CTs were not available for routinely treated patients in our institution, the reason for which we included the first two control CTs for anatomical robust optimization in this study. Although the research interest legitimates this approach as these CTs are closest to the planning CT, it remains a non-ideal solution for two reasons. (1) There was introduced some potential bias in advantage of the aRO plans since the performed robustness analysis considers for the calculation of total cumulative and perturbed dose distributions also the same two CT data sets used for the optimization. (2) It cannot be excluded that some systematic non-rigid anatomical changes have already started within the first 2 weeks of treatment, meaning that aRO plan optimization did not only include random anatomy changes. Supplementary Material 1 contains further investigations indicating that the first two control CTs present mainly random non-rigid positioning variations and are less affected by relevant treatment-induced anatomical changes compared to the last control CT.

When proving whether the findings of the presented study could be reproduced in an actual clinical setting with the proposed protocol, the unfavorable aspects as the additional patient CT dose and the greater effort in the clinical workflow for multiple planning CT acquisition must be weighed against the expected benefit in plan robustness and reduced need of treatment adaptation. Of note, the imaging dose for the additional CT scans can be lowered compared to the scan that is used for target and OAR delineation. In the case that no additional planning CTs were acquired but a need for plan adaptation is identified based on control CT imaging during the treatment course, an anatomical robust optimization could at least be performed during the replanning procedure by considering the latest control CTs for generating an adapted plan with improved robustness.

Our study has additional limitations. First, our retrospective analysis to generate the perturbed dose distributions was intensively time-consuming, lasting up to 9 h for one plan (for 30 cumulative perturbed doses) per patient, due to the required calculation of perturbed doses for each fraction. This would be a serious limitation for a prospective robustness analysis. Second, there are limitations related to the image registration procedure, which are well documented in the literature and may lead to uncertainties in the calculation of cumulative doses.^29,30 And last, we omitted the robustness evaluation on the planning CT(s), since the objective was to evaluate the overall plan robustness simulating a realistic fractionated treatment course.

Conclusions

With a comprehensive multiscenario evaluation, anatomical robust optimization of IMPT plans for HNSCC patients showed superior robustness against anatomy, setup and range uncertainties in comparison with classical robust optimization. Non-rigid anatomical variations during the treatment course play the most critical role in target coverage degradation for cRO plans. The aRO approach preserves the target coverage due to anatomy variations in most of the cases, maintaining at least similar robustness against setup and range uncertainties as the cRO plans. Clinical implementation of aRO is in principle feasible, including additional planning CT data sets in the optimization. Additional dosimetric and organizational aspects must be weighed against the expected benefit in plan robustness and reduced need of plan adaptation.