Dose comparison of proton beam therapy with and without a multileaf collimator in the treatment of multiple hepatocellular carcinoma

Graphical abstract

Highlights

• •Multi-leaf collimators reduced median V_1Gy(RBE) by 2.4%.
• •Liver sparing strongly correlated with inter-lesion distance (r_s = 0.95).
• •Multi-leaf collimators increased shallow doses to skin and ribs by up to 4.8%.

Abstract

Background and purpose

Proton beam therapy (PBT) offers dosimetric advantages over conventional photon-based radiation therapy in treating hepatocellular carcinoma (HCC) because it more effectively spares normal liver tissue. This study investigated the potential benefits of incorporating a multileaf collimator (MLC) into PBT to further reduce the radiation exposure of normal liver tissues.

Materials and methods

We compared eighteen PBT treatment plans with and without MLC implementation, focusing on normal liver dose metrics. Two normal-tissue complication probability (NTCP) models were used to evaluate treatment-related liver side effects. Additionally, we examined the correlation between the characteristics of multiple treatment targets and the volume of liver sparing at low doses to determine the criteria for MLC application.

Results

The V_1Gy(RBE) to the normal liver was reduced by a median of 2.4% (IQR: 0.8–6.0%) in the MLC-based plans. Implementing an MLC significantly reduced treatment-related liver side effects according to the results of the NTCP models. The maximum increase in the shallow-depth dose was 4.8%. The liver-sparing gain at ≤ 1 Gy (RBE) and the cumulative minimum distance between multiple HCC lesions were positively correlated (r_s = 0.95).

Conclusions

The integration of MLC into PBT planning effectively reduced low-dose spread and increased normal liver sparing, potentially lowering the risk of radiation-induced liver side effects. However, the associated increase in the maximum dose at shallow depths warrants careful consideration during plan optimization. The spatial relationship between multiple lesions may serve as a useful criterion for selecting MLC use in clinical settings involving multiple targets.

1. Introduction

Radiotherapy is an important locoregional treatment for hepatocellular carcinoma (HCC), but dose escalation is limited by the radiosensitivity of the hepatic parenchyma and the risk of radiation-induced liver disease (RILD). Proton beam therapy (PBT) offers dosimetric advantages over conventional photon-based radiotherapy by better sparing normal liver tissue from low to moderate radiation doses, which offers clinical benefits for patients with liver tumors [1], [2], [3]. Several clinical studies have demonstrated that, compared with photon therapy, PBT significantly improves overall survival by minimizing treatment-related liver side effects [4], [5], [6], [7], [8], [9].

The multileaf collimator (MLC) is an essential component of modern radiotherapy systems. In the context of PBT, integrating collimator systems with pencil beam scanning (PBS) enhances lateral field-shaping capabilities and may improve dose conformity [10], [11], [12]. The physical rationale is that the lateral penumbra of individual proton spots broadens with depth because of multiple Coulomb scattering, increasing low-dose exposure to surrounding normal tissues. By truncating these low-dose regions, the MLC sharpens field edges and reduces normal tissue irradiation. Prior studies have demonstrated the clinical advantages of PBS with collimation in improving plan quality and normal tissue sparing across various tumor sites, including the maxillary sinus, brain, and lungs [13], [14], [15].

The aim of this study was to evaluate the dosimetric advantages of incorporating an MLC into pencil beam line scanning (PBLS) proton therapy in the treatment of HCC.

2. Materials and methods

2.1. Patient data

This retrospective study included eighteen patients with multiple HCC lesions who underwent PBLS proton therapy at our institution and was approved by the Institutional Review Board of Chang Gung Medical Foundation (IRB No.: 202301813B0). Respiratory motion was managed using one of three mitigation techniques based on the results of respiratory training: abdominal compression (n = 13), respiratory gating (n = 3), or breath holding (n = 2). Four-dimensional computed tomography (4DCT) was used to assess respiratory motion. For patients treated with abdominal compression or respiratory gating, the average intensity projection CT or a specific respiratory-phase image, respectively, was used as the planning CT. For breath-hold patients, three breath-hold CT scans were acquired. Since each scan may exhibit slight baseline variation among breath holds, the dataset with the least target positional deviation relative to the others was selected for treatment planning.

Radiation oncologists defined the gross target volume to cover gross liver tumors, sites of tumor thrombosis, and gross regional lymph nodes. According to the physician’s clinical judgment, an additional margin of 0–5 mm may be added to generate the clinical target volume (CTV) to cover adjacent risky areas. The internal target volume (ITV) was subsequently generated by incorporating motion information from the relevant 4DCT phases.

2.2. Treatment planning

The beam angles were selected by the dosimetrist to minimize the path length through normal liver tissue while avoiding traversal of the gastrointestinal (GI) tract. Each HCC lesion was treated with at least two distinct beam angles, with no single beam contributing more than 70% of the total dose. Collimator angles were chosen to maximize beam conformity from the beam’s eye view. The distributions of the selected beam and collimator angles are presented in Supplementary Fig. S1 and Fig. S2(A), respectively. The carbon-steel MLC consisted of 46 opposing leaf pairs—26 central (3.4 mm) and 20 outer (5.4 mm)—and could operate at collimator angles of 0°, 90°, and 270°. The MLC margins were initially set to 5 mm between the leaf and target edges and were incrementally increased to 8 mm as needed based on the robustness optimization results, ensuring adequate target coverage while minimizing lateral penumbra broadening. When the margin was insufficient, the optimizer compensated by overweighting boundary beamlets, producing peripheral hot spots; in such cases, we increased the MLC margin accordingly. The distribution of the actual used MLC margin is shown in Fig. S3. In this study, a 4.0-cm high-density polyethylene range shifter was used. The beam energies actually applied in the patient plans ranged from 70.4–188.0 MeV, corresponding to spot sizes of 0.42–1.03 cm. The spacing parameters were 1.0σ for energy layers, 0.6σ for lines, and 0.3σ for lateral spots.

All patients received a prescribed dose of 72.6 Gy(RBE)2 in 22 fractions to the ITV, assuming a constant RBE of 1.1 [16]. Robust optimization was performed by simulating setup variations of ± 3.0 mm and range uncertainties of ± 3.5%, applied to the ITV. Dose guidance was applied to normal tissues, including maximum dose (D_max, defined as the dose to a single voxel) limits for the spinal cord, GI tract, and chest wall. Volumetric guidance included V_1Gy(RBE) and V_30Gy(RBE) for the normal liver and V_18Gy(RBE) for the kidney. GI D_max guidance carried the highest penalty weighting, and the dose to the target near normal tissues was allowed to decrease to 60% of the prescribed dose when necessary. Details of the dose guidance for normal tissues are summarized in Supplementary Table S1, in which liver guidance is based on a study by Lee et al. [17].

Treatment plans incorporating the MLC were optimized using a Monte Carlo algorithm (IonMonteCarlo v5.1, RayStation RS10B) with 0.5% statistical uncertainty. To mitigate the interplay effect during PBS delivery, each beam was delivered with at least two repaintings [18], [19]. The distribution of selected repainting numbers is presented in Supplementary Fig. S2(B). For comparison, non-MLC (nMLC) plans were reoptimized using the same parameters but without MLC application. Dose-volume histogram (DVH) metrics were normalized so that the D_100% of the ITV in the nMLC plans matched that of the MLC plans.

For multiple‐target PBLS treatments, the MLC serves two purposes: it sharpens the lateral penumbra and suppresses unintended low-dose irradiation between spatially separated lesions. This second function is particularly relevant in our workflow because the current TPS cannot “beam-off” segments between disjoint targets, which leads to avoidable low-dose spill (Supplementary Fig. S4, orange traces). To decouple this effect from system-specific delivery characteristics and approximate spot-scanning behavior, we introduced a “spot-isolated MLC simulation plan,” in which any field that originally covered more than one lesion was split into sub-fields sharing the same gantry angle but dedicated to a single target in our TPS, with and without an MLC.

2.3. Dose distribution and robustness evaluation

Plan robustness was evaluated by simulating setup variations of ± 3.0 mm independently along the anteroposterior, lateral, and superoinferior axes, resulting in six isocenter shift scenarios, combined with ± 3.5% range uncertainties, for a total of twelve scenarios on the planning image. To evaluate the robustness against respiratory motion–induced deformation of the target and organs, dose recalculations were performed on the maximum inhalation and exhalation phases of 4DCT for patients treated with abdominal compression or respiratory gating. For breath-hold patients, recalculations were based on the two breath-hold CT datasets not used for planning. The worst-case dose was defined as the primary robustness dose metric across all uncertainty scenarios.

To quantify the impact of MLC use on the entrance dose, pixel-by-pixel comparisons were performed to calculate the maximum dose increase in the entrance region of the spread-out Bragg peak (SOBP) across various depths relative to the skin surface by subtracting the nMLC dose from the MLC dose. The corresponding depth-dependent dose–area product is shown in Supplementary Fig. S5.

2.4. Characteristics of multiple treatment targets

The cumulative minimum distance between irradiated targets was calculated to assess the influence of intertarget spacing on normal liver sparing. The cumulative minimum distance, d(%), was defined as follows:

$$d(\%)=(\sum _{i=1}^{n}w(i)d_{\mathit{min}}(i)/T(i))x100\%$$ (1)

where n is the number of treatment fields, d_min(i) is the shortest distance between targets (measured perpendicular to the collimator leaves in the beam’s eye view), T(i) is the total irradiated width encompassing the outer boundaries of the targets in that view, and w(i) is the relative monitor unit weighting for each field (Supplementary Fig. S4A). When the targets were too close to the MLC leaves to effectively block the space, d_min(i) was set to zero (Supplementary Fig. S4B).

2.5. Statistical analysis

The Wilcoxon signed-rank test was used to compare MLC and nMLC plan metrics, including a predicted posttreatment Child–Pugh (CP) score increase ≥ 2 and an albumin-bilirubin (ALBI) grade increase ≥ 1 based on the normal-tissue complication probability (NTCP) model proposed by Pursley et al. [20]. The potential benefit in reducing the risk of RILD was estimated using the ratio of unirradiated liver volume (ULV) to standard liver volume (SLV) (% of ULV/SLV), based on the risk stratification reported by Hsieh et al. [21]. Spearman correlation was employed to assess the relationships between the liver-sparing gain and clinical variables. A generalized estimation equation was used to evaluate the correlations between the increases in entrance dose and various depths from the skin surface.

3. Results

The comparison of the dose distributions of the nMLC and MLC plans revealed that incorporating the MLC resulted in reduced radiation dose to the normal liver while maintaining adequate target coverage, as illustrated in the example of the DVH comparison in Fig. 1C. Specifically, the median reductions in V_1Gy(RBE), V_15Gy(RBE), V_30Gy(RBE), and mean dose to the normal liver were 2.4% (IQR: 0.8–6.0%), 1.3% (0.5–4.0%), 2.0% (0.6–3.3%), and 5.5% (3.6–11.4%), respectively, in the MLC group (Fig. 2).

Fig. 1.

Differences in the DVH and dose distribution between the MLC and nMLC plans in one of our study cases. (A) Dose distribution of the treatment plan in the MLC group. (B) Dose distribution of the treatment plan in the nMLC group. (C) Comparison of the DVH between A (solid line) and B (dotted line) (green lines: normal liver; blue lines: CTV; red lines: ITV). (D) Dose difference between A and B (A-B). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 2.

Comparisons of normal liver doses and volumes between the MLC and nMLC groups. The normal liver volumes in the V_1Gy(RBE), V_15Gy(RBE), and V_30Gy(RBE) groups, as well as the normal liver volumes at the mean dose, were significantly lower in the MLC group than in the nMLC group.

According to the NTCP models, MLC implementation was associated with a lower risk of posttreatment CP score worsening (p = 0.005, mean difference = 0.3%) and ALBI grade worsening (p ≤ 0.001, mean difference = 0.9%). Furthermore, compared with the nMLC plan, the MLC technique achieved a median gain of 2.2% (IQR: 0.1–5.4%) in the ULV/SLV ratio. Notably, this improvement in volume preservation translated to a decreased estimated risk of RILD in 22.0% of the patients (Fig. 3).

Fig. 3.

Estimated risk of radiation-induced liver disease. Data are based on the Eastern cohort from Hsieh et al. [20]. ULV indicates unirradiated liver volume, defined as absolute liver volume receiving less than 1 Gy(RBE); SLV, standard liver volume computed using the Urata formula.

An example of the spatial dose distribution differences between the MLC and nMLC plans is shown in Fig. 1D. MLC use resulted in sharper dose gradients at the edges of the target and in the normal liver regions between adjacent targets. However, the MLC plans also resulted in a modest increase in the entrance dose within the SOBP and a slight increase in the maximum dose within the treatment target. A positive correlation (p = 0.003) was observed between the magnitude of the dose increase and depth from the skin surface in the MLC plans, with the greatest increase in dose occurring at shallower depths (Fig. 4). In our cohort, the maximum increase in the entrance dose was 4.8% of the prescribed dose at a depth of 0–1 cm from the skin surface when an MLC was used.

Fig. 4.

Dose variations between the MLC and nMLC groups within the entrance area of the SOBP. The maximum increase in dose at various depths relative to the skin is shown in this figure, and the maximum dose increase was calculated pixel-by-pixel by subtracting the dose in the nMLC group from the dose in the MLC group.

In evaluating the impact of MLC use on plan robustness, a statistically significant negative shift in CTV D_100% coverage occurred in only two of the simulated scenarios, with maximum decreases of 3.3% and 4.0%, while the MLC group consistently demonstrated reduced normal-liver dose across all robustness simulations, mirroring the improvements observed in the nominal plans.

Spearman analysis demonstrated a strong relationship between the liver-sparing gain at ≤ 1 Gy (RBE) and the d(%) between multiple HCC targets (r_s = 0.95, p < 0.001) (Fig. 5).

Fig. 5.

Correlation analysis between the V_1Gy(RBE) liver-sparing gain and the sum of the minimum distance between multiple HCC lesions. The Spearman correlation coefficient was 0.95, with a significant p value (p < 0.001), suggesting a strong correlation between the liver-sparing gain and the distance among treatment targets.

In the spot-isolated MLC simulation plan, although the benefit of MLCs was attenuated by removing their ability to suppress unintended low-dose irradiation between spatially separated lesions, the MLC configuration still significantly reduced V_1Gy(RBE), V_15Gy(RBE), V_30Gy(RBE), and mean dose to the normal liver. The median reductions were 1.2% (IQR 0.0–4.1%), 1.8% (0.5–4.1%), 2.3% (0.6–4.1%), and 5.9% (3.7–10.7%), respectively (Fig. 6).

Fig. 6.

Comparisons of normal liver doses and volumes between the MLC and nMLC groups in the spot-isolated MLC simulation plan. Significant reductions in the dose delivered to the normal liver were observed when the MLC was used.

4. Discussion

In this study, we evaluated plan robustness from three perspectives: proton range uncertainty, setup variation, and respiratory motion. Our results showed that MLC implementation lowered low-dose liver exposure, decreasing V_1Gy(RBE) by up to 23.3% and leading to a reduction in the estimated RILD risk for 22.0% of the patient cohort. Furthermore, MLC use was associated with lower predicted risks of posttreatment worsening in CP score and ALBI grade.

Nonetheless, the application of an MLC may impact plan robustness, particularly when inappropriate MLC margins are used [11]. A marginal reduction in CTV D_100% was observed in two of the robustness scenarios with MLC use in this study; however, all the reductions were within 0.7% and were deemed clinically acceptable. In cases where clinically significant degradation in target coverage occurs, adjusting the MLC margins may be necessary. In this context, systematic approaches such as modeling the maximum pencil beam center shift have been proposed to guide aperture margin selection in collimated spot-scanning proton therapy [22], providing a more rigorous method for balancing plan quality and robustness without extensive iterative testing.

Due to the heterogeneous use of motion-management strategies in our cohort, this study was not designed to determine which specific technique yields the greatest benefit from MLC implementation. Importantly, respiratory motion was consistently incorporated into the ITV, MLC margins were expanded beyond the ITV, and multiple repaintings were applied in all cases. Under these common planning conditions, we posit that the residual impact of the specific motion-management technique on MLC robustness is likely to be small.

While 4D robust optimization accounts for anatomical variations across respiratory phases, it does not explicitly model the temporal delivery sequence of PBS [23]. Prior studies have shown that adding repainting or incorporating time-structure uncertainties into the optimization process can further reduce interplay effects [24], [25]. Although the time-dependent delivery sequence was not explicitly modeled in this study, each beam was delivered with at least two repaintings to mitigate potential interplay effects, a strategy previously shown to effectively reduce motion-induced dose inhomogeneities in liver proton therapy [19], [26], [27].

Previous studies have reported that MLCs can contribute to a 3–25% increase in the entrance dose at shallow depths because of collimator-scattered protons and edge enhancement in PBS proton therapy, with both physical and biological effects observed [28], [29]. This mechanism is consistent with our observation of an increased shallow-depth dose in the MLC plans; Fig. 4 highlights how MLC use affects dose distributions at various depths. While this increase was modest, it may have clinical implications, particularly for normal tissues near the skin, such as the ribs and the skin itself. This is especially relevant in hypofractionated proton therapy, where even small dose increases may have a greater biological impact [30].

In addition to prior findings showing that MLCs reduce the absolute field penumbra in single-target plans, especially for shallow targets, small volumes, and large air gaps [28], our spot-isolated MLC simulation plan similarly demonstrated greater relative V_1Gy(RBE) sparing for smaller targets (Supplementary Fig. S6). This occurred because the physical penumbra width, primarily determined by spot size and scattering, remains nearly constant and therefore represents a larger fraction of the field in smaller targets. Moreover, aperture-scattered protons caused more pronounced shallow-depth dose escalation for smaller openings, further enhancing the dose-shaping effect. Furthermore, our study identified a novel parameter relevant for multitarget cases, d(%) (Equation (1). When the overall irradiated width was similar across beams, a larger minimum distance between targets indicated better geometric separation, allowing the MLC to block more normal liver between lesions. Conversely, when the minimum distance between targets remained the same, a wider overall irradiated field reduced liver sparing because it typically corresponds to (1) decreased conformity arising from collimating based on the maximal lateral extent of the target in the beam’s eye view and (2) target positioning near the liver boundary, where less normal liver can be effectively shielded by the MLC.

Recent preclinical and early clinical data suggest that low-dose radiation may reprogram the tumor microenvironment and enhance immunotherapy responses—the “radscopal effect” [31]. Conversely, clinical evidence indicates that even very low liver doses may worsen hepatic function in patients with impaired baseline reserve [32]. Although an evaluation of combination systemic therapy was outside the scope of this study, whether low-dose exposure influences outcomes when paired with immunotherapy remains an important question to be addressed in future investigations.

Several limitations should be acknowledged. The small sample size may limit the generalizability of our findings. In our TPS, the dose calculation algorithms do not fully account for interleaf leakage or transmission through closed leaf tips. Our measurements revealed that the interleaf leakage and intraleaf transmission were less than 0.87% and 0.31% of the prescribed dose, respectively. These contributions are minor and therefore unlikely to have significant dosimetric effects. Moreover, all treatment plans in our center undergo rigorous patient-specific quality assurance before clinical delivery. Additionally, secondary neutron doses generated by interactions between incident protons and MLC materials were not evaluated. Previous Monte Carlo studies have shown that dynamic or patient-specific collimation in PBS slightly increases secondary neutron production, with the corresponding lifetime attributable risk estimated at approximately 0.18–0.62% for dynamically collimated systems [33] and up to 0.19% for patient-specific apertures [34]. These findings suggest that the additional neutron contribution from the static carbon-steel MLCs used in our study is expected to be negligible; however, long-term follow-up remains warranted to monitor potential late effects associated with neutron exposure. In our institutional guidelines, worst-case dose quantifies the reported endpoints. However, robustness metrics remain method-dependent, and voxel-wise minimum and maximum dose distributions also provide important information. With no internationally accepted standard, the ESTRO Delphi consensus calls for harmonized robustness evaluations and provides an expert-based framework for future practice [35]. Finally, we adopted the NTCP model proposed by Pursley et al. [20], as it is the only model incorporating a substantial proportion of proton-treated patients. Although MLC implementation in our study significantly reduced the predicted liver side effects, the absolute NTCP differences were small—likely reflecting model parameters derived largely from photon datasets, which are less sensitive to low-dose variations. Accordingly, these estimates should be interpreted with caution.

In conclusion, this study demonstrated that compared with non-MLC approaches, MLC-based proton therapy effectively spared a larger volume of normal liver. These findings suggested that MLC use may have reduced the risk of liver side effects and improved clinical outcomes in patients with multiple HCC lesions undergoing proton therapy.

Declarations

Ethics approval and consent to participate

All patient data used in this study were approved for utilization by the Chang Gung Memorial Hospital Institutional Review Board under study #202301813B0.

Declaration of generative AI in scientific writing

The authors declare that no generative AI or AI-assisted technologies were used in the writing or editing of this manuscript.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary material

The following are the Supplementary data to this article: