Skip to main content

The role of a knowledge based dose–volume histogram predictive model in the optimisation of intensity-modulated proton plans for hepatocellular carcinoma patients

Training and validation of a novel commercial system



To investigate the performance of a knowledge-based RapidPlan, for optimisation of intensity-modulated proton therapy (IMPT) plans applied to hepatocellular cancer (HCC) patients.


A cohort of 65 patients was retrospectively selected: 50 were used to “train” the model, while the remaining 15 provided independent validation. The performance of the RapidPlan model was benchmarked against manual optimisation and was also compared to volumetric modulated arc therapy (RapidArc) photon plans. A subanalysis appraised the performance of the RapidPlan model applied to patients with lesions ≤300 cm3 or larger. Quantitative assessment was based on several metrics derived from the constraints of the NRG-GI003 clinical trial.


There was an equivalence between manual plans and RapidPlan-optimised IMPT plans, which outperformed the RapidArc plans. The planning dose–volume objectives were met on average for all structures except for D0.5cm3 ≤30 Gy in the bowels. Limiting the results to the class-solution proton plans (all values in Gy), the data for manual plans vs RapidPlan-based IMPT plans, respectively, showed the following: D99% to the target of 47.5 ± 1.4 vs 47.2 ± 1.2; for organs at risk, the mean dose to the healthy liver was 6.7 ± 3.6 vs 6.7 ± 3.7; the mean dose to the kidneys was 0.2 ± 0.5 vs 0.1 ± 0.2; D0.5cm3 for the bowels was 33.4 ± 16.4 vs 30.2 ± 16.0; for the stomach was 17.9 ± 19.9 vs 14.9 ± 18.8; for the oesophagus was 17.9 ± 15.1 vs 14.9 ± 13.9; for the spinal cord was 0.5 ± 1.6 vs 0.2 ± 0.7. The model performed similarly for cases with small or large lesions.


A knowledge-based RapidPlan model was trained and validated for IMPT. The results demonstrate that RapidPlan can be trained adequately for IMPT in HCC. The quality of the RapidPlan-based plans is at least equivalent compared to what is achievable with manual planning. RapidPlan also confirmed the potential to optimise the quality of the proton therapy results, thus reducing the impact of operator planning skills on patient results.


Hepatocellular carcinoma (HCC) is one of the most common primary liver tumours and a primary source of cancer-related deaths [1]. Its management requires multidisciplinary efforts, and several alternative treatment options are available, from surgery to ablative therapies. Radiotherapy is part of the treatment protocol according to several national and international guidelines (e.g. the European Association for the Study of the Liver and European Organization for Research and Treatment of Cancer (EASL-EORTC) [2]). However, the application of radiotherapy is still limited due to the risk of radiation-induced liver disease (RILD) [3]. The advent of intensity-modulated radiation therapy (IMRT) and volumetric modulated arc therapy (VMAT) opened new options for the implementation of photon radiotherapy for HCC. Wang et al. [4, 5] revealed the beneficial role of VMAT as a treatment technique in a cohort of 138 patients with inoperable HCC, with about 90% of the patients in the American Joint Committee on Cancer (AJCC) stage III or IV, and with tumour volumes ranging from 28 to 3620 cm3. The 1‑year survival rate was 40% (stage dependent), with a median survival of 10.3 months and local control in 94% of the cases. RILD, in this compromised patient population, was observed in 25% of the cases while grade 3 gastrointestinal toxicity was observed in only 12.3% of the patients.

While randomised, controlled trials will continue to clarify the role of proton therapy in the management of HCC [6], this therapy may enable improved control rates with reduced toxicity levels. For example, the Consensus Report from the Miami Liver Proton Therapy Conference in 2018 [7] evaluated the optimal utilisation of proton therapy. The report concluded that proton therapy should be the preferred option when photons would deliver suboptimal therapeutic doses. Sanford et al. [8] retrospectively studied 133 patients who received stereotactic body radiation therapy (SBRT) using either protons or photons with curative intent. With a median follow-up of 14 months, proton therapy was associated with more prolonged survival (median 31 vs 14 months with photons) and a 2-year overall survival of 59.1% for protons versus 28.6% for photons. Patients treated with protons also showed a decreased incidence of RILD.

The main challenge of SBRT in HCC is adherence to dose–volume constraints on organs at risk (OARs) in proximity to the target. Kim et al. [9] reviewed a cohort of 243 patients who received a risk-adapted moderate SBRT proton treatment (in 10 fractions) and concluded that protons have the potential to play a beneficial role at all stages of HCC. In an earlier study investigating the role of intensity-modulated proton therapy (IMPT) for ablative SBRT, we showed the feasibility of IMPT with robust plan optimisation in a cohort of 20 patients with advanced HCC (prescription of 60 Gy in 3 fractions) [10].

The process of optimising the dose distribution in treatment planning requires the orchestrated use of information from several different sources of “knowledge” (clinical, technical, dosimetric) which together determine the most appropriate technique and dose distribution for each patient. From a mathematical standpoint, the proper dose–volume distribution could be modelled by using dose and anatomy data from existing plans to predict an optimal dose distribution for any new patients, based on the specific location and contours of the target tumour. Among the various machine-learning (ML) methods, the RapidPlan approach (RP, Varian Medical Systems, Palo Alto, CA, USA) was developed based on the original work of Washington and Duke Universities [11,12,13,14,15]. It was extensively used for several clinical indications, both in feasibility planning studies and in clinical applications; Ge et al. recently published an extensive review of the data-driven approaches to ML-based planning for IMRT [16]. In summary, RapidPlan is an engine which allows the radiation-oncology team to build and train a dose–volume histogram (DVH) prediction model from a library of preselected high-quality treatment plans. The predictive model, based on a geometric and dosimetric parameterisation of the expected dose on “new” patients, generates dose–volume constraints usable within the inverse planning phase. Fogliata et al. [17] investigated RapidPlan for HCC patients. We developed a predictive model based on a cohort of 45 patients planned for VMAT via RapidArc (RA); another 25 cases provided independent validation. Quantitative equivalence between the clinical and the RapidPlan-based plans suggested the applicability of RapidPlan to a wide spectrum of cases and target volumes. More recently, Yu et al. [18] benchmarked RapidPlan-based against clinical plans for IMRT treatment of liver patients and appraised the relative merits of a “general” versus a “specific” model (based on uniform anatomical characteristics). In this study, the authors showed how high-quality plans were generated with RapidPlan and that the use of specific models could further improve the degree of conformality and spare healthy tissue.

The application of ML-driven planning to proton therapy is still in its infancy, and more evidence-based data is needed. However, Delaney et al. [19] published some pioneering work describing the use of a prototype of RapidPlan for protons and applied it to a head and neck cancer case aiming to develop a strategy to identify the patients best suited for proton therapy. In a subsequent study, the same group demonstrated how the RapidPlan-based plans were consistently comparable to manual plans for IMPT, suggesting the feasibility of using RapidPlan as a tool for the automation of proton planning [20]. ML-based planning for protons is not assumed to be more challenging than for photons but, given its novelty and the different implementations compared to photons, it requires validation in a variety of clinical indications to prove its reliability.

The present study aimed to provide further evidence about applying the RapidPlan automation process to protons in challenging clinical cases. This aim was realised by training and validating a general-purpose model for IMPT for advanced HCC patients (according to the specifications of an ongoing randomised clinical trial) and by benchmarking the performance of this model against that of a published one for photons [17].

Materials and methods

Patient selection, contouring, and dose prescription

The computed tomography (CT) images and structure sets of 65 patients (extracted from the database of 138 cases published by Wang et al. [4, 5]) were used for this in-silico investigation. All patients initially signed authorisation forms allowing the use of their data for further research. The cohort of patients included Barcelona Clinic Liver Cancer (BCLC) stages A–C and Child–Pugh classes A–B in patients who had single lesions >5.0 cm or multinodular lesions >3 cm. All were inoperable and ineligible for other ablative approaches. The gross tumour volume (GTV) was defined as the primary tumour plus abnormal portal areas, and the clinical target volume (CTV) equal to the planning target volume (PTV) was defined as the GTV plus a 1.5–2.0 cm margin. The Organs at RIks (OARs) available for treatment plan optimisation were the healthy liver (defined as the liver minus the CTV), the oesophagus, stomach, bowel bag, kidneys, heart, duodenum, chest wall, and spinal cord.

In the original study that supplied this patient pool [4, 5], 1.8 or 2.0 Gy per fraction were used. The present study applied a hypofractionated scheme of 5 fractions of 10.0 Gy each, as derived from the guidelines of the NRG-GI003 trial (NCT03186898 at, a 17-centre study which is examining proton vs photon therapy for HCC. The OAR dose–volume constraints were defined according to the same NRG-GI003 protocol. The dose coverage required for the PTV was D95% ≥95.0% (90.0% as an acceptable variation) with Dx representing the minimum dose that covers an x fraction of volume (in % or cm3). The near-to-maximum dose constraint to the target was set as D0.3cm3 ≤120.0%.

For the structure defined as the liver minus the target volume (liver-target), the mean dose was Dmean ≤13.0 Gy. For the stomach, bowels, duodenum and oesophagus, the near-to-maximum dose constraint was D0.5cm3 ≤30.0 Gy. For the spinal cord, this dose constraint was 25.0 Gy; for the heart, it was 20 Gy. For the chest wall, the limit was: D2cm3 ≤35 Gy. For the combined and single kidneys, the acceptable mean dose was Dmean ≤10.0 Gy in addition to D33% ≤15.0 Gy and D10% ≤7.0 Gy. Dose constraints to the OARs took precedence over target volume.

For each patient, several sets of plans were designed for photon VMAT RapidArc, manual IMPT planning, and IMPT using RapidPlan according to the techniques described below. All the plans were re-normalised to the mean dose of the PTV for comparison reasons to avoid bias in data representation. The Eclipse treatment planning system (Varian Medical Systems, Palo Alto, CA, USA) with the clinically released version 16.0 was used for all the patients and algorithms.

Photon planning

New photon treatment plans (with RapidArc VMAT) were designed for the 65 patients using a TrueBeam linear accelerator (Varian Medical Systems, Palo Alto, CA, USA) equipped with a Millennimum multileaf collimator with a resolution of 5.0 mm at isocenter using 10 MV flattening filter-free photon beams. The optimisation was performed with the Photon Optimiser algorithm. Two partial arcs were used for all plans with a case-by-case setting of the start/stop gantry angles, collimator angle, and jaw settings to account for the specificity of each target. The final dose calculation was performed employing the Acuros-XB engine with a grid of 2.5 mm [21].

Proton planning

The IMPT plans were created using pencil beam spot scanning from the ProBeam proton system (Varian Medical Systems, Palo Alto, CA, USA). The reference beam data in Eclipse, derived from the commissioning of the Scripps Proton Center (San Diego, CA, USA) were used with permission. The dose distribution was optimised using the fluence-based nonlinear universal Proton Optimizer (NUPO) [22]. Spot spacing was 0.425 of the energy-dependent in-air full-width half-maximum spot size at the isocenter. The multifield simultaneous spot optimisation method was selected for all plans. A range shifter was used for all fields. The Proton Convolution Superposition algorithm was used for the final dose calculation with a grid of 2.5 × 2.5 × 2.5 mm3. A constant relative biological effectiveness (RBE) of 1.1 was applied.

Robust optimisation was enabled for the target along the craniocaudal axis with a positioning uncertainty of 3 mm.

All patients were planned with two different beam settings:

  • An individualised beam arrangement with 2–3 fields whose entrance angles were tuned according to the target position (IMPT (a));

  • A class solution using two beams with gantry angles of 20° and 300° for anterior targets and 290° and 210° for posterior targets (IMPT (b)).

The beam angles for groups (a) and (b) were modestly different, with the gantry angles variations somewhat limited due to dosimetric constraints on the various OARs. Both groups of plans were optimised twice: once with a “manual” trial-and-error procedure and once using the RapidPlan model.

The photon and proton RapidPlan predictive models

The photon RapidPlan model discussed in Fogliata et al. [17] was used as a starting point for optimisation of the VMAT RapidArc plans. The dose constraints in the model were adjusted per the NRG-GI003 protocol. No further adjustments requiring any new training were applied. The 45 patients used for the photon RapidPlan model were also included in the current study.

A dedicated RapidPlan model for the IMPT plans was defined and trained. The RapidPlan performance is the focus of the present investigation, with manual and RapidPlan-based proton plans (IMPT protons, IMPT_RP) optimised and comparatively analysed. While RapidPlan for photons considers volume partitioning into four subregions—(1) the out-of-field region, (2) the leaf-transmission region, (3) the in-field region and (4) the target overlap region—the proton RapidPlan partitioning considers only subregions (1), (3) and (4). For this reason, a photon RapidPlan model cannot be used for protons. Delaney et al. [20] showed the fundamental difference in the volume partitioning in RapidPlan for photons and protons; readers can refer to that original publication for further details.

For training the proton RapidPlan model in this study, a subset of 50 cases was selected from the class-solution group (b) of plans. The upper and line-type model generated objectives and priorities were defined in the predictive model and then used to define prospective per-patient optimisation objectives as described in Table 1. Line-type objectives are available in RapidPlan-based optimisation to limit the dose in a given structure for all volume levels with the same “strength” in each region. Although robust optimisation is essential in clinical practice, it is strongly advisable but not mandatory for creating a RapidPlan model. Only the geometrical relations between anatomical structures and dose distributions are considered in training, irrespectively from how the dose maps were generated.

Table 1 PTV and OARs objectives implementation in the RapidPlan model for protons

The critical element of the RapidPlan training phase consists of the parametrisation of the dosimetry and geometry data from the training plans. This is done by means of a principal component analysis (PCA) of the in-field dose. The PCA is a transformation that redefines the data in a new coordinate system constructed by “principal components” (PC) such that the greatest variance of the data lies on the first coordinate of the new system (called first principal component). Referring to the DVH, the (first) PC is a curve, having the same presentation (number of data points, normalisation, etc.) as the DVH. It is used to describe the variability within the given set of DVH curves from training. The assumption is that it is possible to characterise any given DVH as the sum of the mean DVH plus a series of PCs multiplied by appropriate coefficients. In practice, the first PC is determined by first calculating the mean DVH for the training set. Next, the mean DVH is subtracted separately from each DVH curve. The remaining residual curves define the full variability of the original set. The first PC is determined by requiring that each residual curve can be represented as the sum of the first PC (multiplied by its coefficient) and the next order residual curves. The first PC, therefore, describes the general variance of the population of DVH curves, but it does not adequately describe more detailed shapes of the curve. For that, it is necessary to include the higher-order components. Any consecutive PC is chosen with the expectation that 2–3 PCs should be sufficient to account for at least 95% of the variance. The specific form of each PC in RapidPlan is not accessible to the users, which might be considered as a limit of the system. An explanatory diagram is presented in the supplemenary materials (Fig. 1Supp). The model so achieved can be used then to predict a probable DVH for any prospective patient and any OAR included in the model itself.

When a model is trained, an internal 10-fold cross-validation is performed. The training set divided into ten parts, and ten models are trained. For each training, a different tenth of the data points (plans) is omitted and used for validating the model. Several standard statistical tools are incorporated in RapidPlan to appraise model quality. A well-trained model should be able to predict data which precisely and accurately reproduce the original training set. A measure of the goodness of the prediction is given by the regression coefficient χ2 (based on the Pearson’s χ2, it is named as the regression model’s parameters average χ2 in the RP environment, and it is derived from the residual difference between the original data and the estimates). Numerically, the closer the regression coefficient χ2 is to one, the better the quality of the model. A further measure is given by the mean squared error of the residuals. A valuable qualitative tool is provided by the scatter plots with the regression fit of the estimated (abscissa) versus achieved (from the training set) of the first principal component (i.e. the residual plots). The scatter plot is the most valuable (qualitative) evaluation of the model and enables quick identification of outliers. In this study, we chose not to remove any of the potential outliers from the model. The average of the validation statistics generated during these 10 model training rounds is then shown in the training log.

Validation experiments

The remaining 15 patients, not used for the proton model training, were selected for independent validation tests (open-loop validation). The comparison was performed on a set of 5 plans per patient:

  • A photon RapidArc plan optimised by using the RapidPlan model published previously [17] with the dose–volume model objectives adapted to the NRG-GI003 protocol.

  • Four IMPT plans with manual and RapidPlan-based optimisation for the beam arrangements of groups (a) and (b).

As an added appraisal of the proton RapidPlan model’s performance, all patients were stratified into two subgroups according to the target volume: PTV ≤300.0 cm3 and PTV >300.0 cm3 (presented in the supplementary materials).

Quantitative assessment of dose–volume metrics

Quantitative metrics were derived from the DVH and included the mean dose and a variety of Dx and Vx parameters, with Vx representing the volume receiving at least an x level of dose (in % or Gy) according to the NRG-GI003 protocol. All parameters could be expressed either in absolute (Gy or cm3) or relative (%) terms. The average DVHs were computed, for each structure and each cohort, with a dose binning resolution of 0.02 Gy. Proton doses are reported in Cobalt Gray equivalent (corrected for the RBE factor).

The Wilcoxon matched-pairs signed-rank test was applied to evaluate the significance of the observed differences per each couple of plans. The threshold for statistical significance was p < 0.05.

Determining the time efficiency of the RapidPlan approach compared to manual planning was not a primary objective of the study since it might be strongly affected by logistics (hardware of the treatment planning system) and subjective factors (planners’ skills). A relative appraisal of the efficiency aspects will be qualitatively presented.


Fig. 1 shows the scatter plot of the first principal component in the model versus the corresponding estimated one for some of the OARs. A compact (narrow) distribution correlates to a good quality of the model’s predictive power. The total mean model estimated quality (the mean squared error of the residuals) was 0.01 ± 0.01. The mean regression mode’s parameter average χ2 was 1.16 ± 0.15.

Fig. 1

Scatter plot and regression lines for some of the various principal-component analysis as a result of the training internal validation tests. DVH dose–volume histogram

Fig. 1 suggests the potential presence of two outliers in the training set, for whom the target volume was located in the proximity of the stomach, leading to higher irradiation compared to the other patients. These two cases correspond to the upper lines in the distribution of the DVH (Fig. 2 in the supplementary materials). For those patients, the target volume was located in the proximity of the stomach and lead to relatively higher irradiation compared to the other patients. Therefore, the outlier status was not considered a valid reason for exclusion from the training set. Instead, it might suggest the need to increase the number of cases to better account for interpatient variance. On the other hand, for proton plans, only a few patients would present such a significant organ involvement.

Fig. 2

Scatter plot of the predicted versus achieved mean doses for the liver-CTV, kidneys, bowel bag, oesophagus and stomach (a) and the corresponding plot for the near-to-maximum D1% for the oesophagus, bowel bag, spinal cord and stomach (b). The graphs include both training (circles) and validation (triangles) sets. CTV clinical target volume

To further appraise the predictive power of the model, we compared the predicted and achieved mean and near-to-maximum doses for some OARs after full optimisation and dose computation. Fig. 2 shows the scatter plot of the mean doses for the liver-CTV, kidneys, bowel bag, oesophagus and stomach and the corresponding plot for the near-to-maximum D1% for the bowel bag, oesophagus, spinal cord and stomach. The graphs include both training and validation sets. The linear regression fit parameters, with a regression coefficient ranging from 0.98 to 0.99, demonstrated how precisely the predicted doses correspond to the achieved values.

Fig. 3 shows the average DVHs for all of the five plans computed in the validation cohort for the target volume, healthy liver (liver minus the PTV), bowel bag, oesophagus, stomach, and healthy tissue (body minus the PTV). The corresponding graphs for all the other OARs are shown in Fig. 3Suppl. The qualitative results drawn from these graphs are the following: (1) the IMPT plans, either with manual planning or with RapidPlan, outperformed the RapidArc photon plans; and (2) the IMPT RapidPlan-based plans are somewhat preferable to the manual plans for certain OARs (e.g. the bowel bag). In assessing beam arrangement, the choice of a class solution or an individualised determination of the angles did not lead to macroscopic differences for most of the OARs (except the spinal cord, where a class solution was visually better).

Fig. 3

Average dose–volume histograms for the entire validation cohort for all the techniques. RA RapidArc, IMTP Intensity modulated proton therapy, RP RapidPlan, PTV planning target volume

The quantitative analysis of the individual DVHs from all of the plans is summarised in Table 2. No clinically meaningful differences were observed among the techniques.

Table 2 Summary of the quantitative analysis on the target volume and organs at risk. Only the validation patients are included

Concerning the OARs: first, the data confirmed that IMPT outperformed RapidArc in all metrics. Second, the IMPT (a) or (b) plans performed equivalently with no statistically significant difference for almost all parameters. Third, there is a general trend for RapidPlan plans to be superior to manual plans, although the differences are small.

All the techniques did not respect the constraints to D0.5cm3 for the bowels while the D1cm3 resulted violated in ≤30.0 Gy for all techniques. Alternative constraints might be used in different protocols, e.g. we reported V36Gy frequently thresholded at 3.0 cm3 and, in this case, all techniques met the requirements.

Comparisons of the IMPT_RP (b) plans for small and large lesion subgroups and the results are reported in the supplementary materials.

RapidPlan resulted in an average time saving of about 20–30% compared to manual planning, mostly because RapidPlan optimisation can proceed unattended by staff. With a conventional manual approach, even if aided by a well-conceived dose–volume objectives template, each patient required some degree of adaptation of constraints as well as trial-and-error attempts to properly manage the various trade-offs between conflicting planning aims.


This report provides results of an implementation study of the RapidPlan knowledge-based approach to IMPT planning for HCC. In this respect, it aims to expand the usability of the RapidPlan methods, typically used with photon therapy, to include proton therapy as well. Since the potential role of IMPT for HCC has been investigated at the planning and clinical levels by several groups [6, 7, 9, 10], it was, therefore, natural to begin assessing whether RapidPlan could aid in IMPT treatment planning. The results in this report demonstrate that RapidPlan models can be trained adequately for IMPT to target HCC and that the quality of the RapidPlan-based plans are equivalent to, and in some instances are better than, manual planning.

Our results are consistent with the earlier findings of Delaney et al. [19, 20] and demonstrate the versatility of RapidPlan for proton therapy. We offer several high-level conclusions about our results. First, the size of the cohort needed to train a reliable IMPT model can be moderately small; 50 patients were, in fact, sufficient to train this model (Fig. 2). Second, the RP based plans also resulted in substantially equivalent with respect to the choice of the beam entrances. These results, valid for the present study and to be further proven in other anatomical districts (depending on the geometrical features of the targeted anatomy), confirm that observed for photons [17]. Still, it is quite relevant since, in protons, the plan quality dependence from the beam orientation is stronger than with photons. The conclusions might be different for cases in which there is much more freedom to select beam entry angles.

Third, the strong correlation between model predictions and achieved dose–volume results suggests the feasibility of automated methods for selecting candidates for proton therapy. Robust and reliable algorithms are needed to establish selection criteria for appropriate use of such labour- and cost-intensive technology, as already outlined in some investigations [23,24,25,26,27,28]. We anticipate that the combination of DVH prediction methods with normal tissue complication probability or secondary cancer induction estimation models would permit a proper selection. Specific tools and guidelines may be elucidated in a future study.

Fourth, although addressed at a mostly qualitative level, the data suggest a relevant time-sparing effect when using RapidPlan vs manual methods, with no compromise in plan quality. Use of RapidPlan may allow clinicians to reserve time for the more complex cases and the refinement of treatment strategies.

Fifth, while it is not yet clear whether RapidPlan models should be trained as broad-scope or narrow-scope tools, Yu et al. [18] in their photon study suggested different levels of sparing OARs or normal liver when using special or general models. Our data indicate that broad-scope proton models can properly manage a large variance in the target volumes (which was similarly shown with photons [17]). This was also suggested by the results from the complementary study presented in the supplementary materials.

It is important to point out that the normalisation method applied in the present study (the mean dose to the PTV) differs from the International Commission of Radiological Units report 83 [29] recommendation to use the median dose. We selected the mean dose for consistency with earlier published studies [5, 17] and also noted the minimal difference in mean vs median values. The average median dose values for the IMPT plans (type B) resulted in 49.54 ± 0.06 Gy, while for the RapidArc plans resulted in 50.03 ± 0.05 Gy with a difference of 1% for IMPT and <0.1% for RapidARc.

A potentially limiting factor in the entire implementation of the RapidPlan in Eclipse is that final users have minimal access to its core, the principal component analysis. This is reasonably unavoidable for any type of commercial solution but, as a good practice recommendation, it would require accurate validation experiments before clinical implementation.

Another potential study limitation is the small number of cases (n = 50) and the associated limited sampling of OAR volumes. Before implementing RapidPlan models, clinicians might need to initially incorporate a more extensive selection of cases to train the model—which is also an element of our future research. Nevertheless, as outlined above, the strong correlation between predicted mean and near-to-maximum doses suggests that a modest cohort can be sufficient for modelling.

Concerning the segmentation of the OARs in the abdomen, many structures (as the duodenum or the bowels) are challenging. In particular, the need to segment the entire bowel bag or the individual loops remains debated. We advocate employing the safest approach, which is the segmentation of the whole bag.

A challenge in HCC treatment is the motion of internal OARs due to respiration and the peristaltic motion of the gastrointestinal tract. The present study analysed only the generation and training of a RapidPlan model and not motion mitigation. Appropriate 4D delivery methods, such as breath-hold or gating [30], could mitigate most of the effects of respiration. More advanced approaches, including rescanning or repainting, could offset the possible interplay of motion with the proton scanning patterns [31,32,33,34,35,36,37,38,39]. Moreover, the use of robust optimisation methods, applied to the CTV (where both range and positioning uncertainties can be modelled into the optimisation process) can reduce the risks associated with internal organ motion.

Finally, it is important to note that robustly optimised plans can be used as the input set for RapidPlan model training and it is an advisable good practice (although not technically mandatory).


A knowledge-based RapidPlan model was trained and validated for intensity-modulated proton therapy (IMPT). The results demonstrate that RapidPlan can create models for effective IMPT delivery in patients with hepatocellular cancer (HCC). The quality of the RapidPlan-based model is at least equivalent to what is achievable with manual planning. In addition, RapidPlan for IMPT shows the potential to optimise results, thus mitigating the impact of planner skills on outcomes. The use of knowledge-based methods can also improve the efficiency of the treatment planning process without compromising the quality of the results.


  1. 1.

    Bray F, Ferlay J, Soerjomataram I, Siegel RL, Torre L, Jemal A (2018) Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 68:394–424

    Google Scholar 

  2. 2.

    European Association for the Study of the Liver, European Organisation for Research and Treatment of Cancer (2012) EASL-EORTC clinical practice guidelines: management of hepatocellular carcinoma. J Hepatol 56:908–943

    Article  Google Scholar 

  3. 3.

    Dawson LA, Normolle D, Balter JM, McGinn CJ, Lawrence TS, Ten Haken RK (2002) Analysis of radiation induced liver disease using the Lyman NTCP model. Int J Radiat Oncol Biol Phys 53:810–821

    Article  Google Scholar 

  4. 4.

    Wang PM, Hsu WC, Chung NN, Chang FL, Fogliata A, Cozzi L (2013) Radiotherapy with volumetric modulated arc therapy for hepatocellular carcinoma patients ineligible for surgery or ablative treatments. Strahlenther Onkol 189:301–307

    Article  Google Scholar 

  5. 5.

    Wang PM, Hsu WC, Chung NN, Chang FL, Fogliata A, Cozzi L (2012) Radiation treatment with volumetric modulated arc therapy of hepatocellular carcinoma patients. Early clinical outcome and toxicity profile from a retrospective analysis of 138 patients. Radiat Oncol 7:207

    Article  Google Scholar 

  6. 6.

    Yeung RH, Chapman TR, Bowen SR, Apisarnthanarax S (2017) Proton beam therapy for hepatocellular carcinoma. Expert Rev Anticancer Ther 17:911–924

    CAS  Article  Google Scholar 

  7. 7.

    Chuong MD, Kaiser A, Khan F, Parikh P, Ben-Josef E, Crane C et al (2019) Consensus report from the miami liver proton therapy conference. Front Oncol 9:457

    Article  Google Scholar 

  8. 8.

    Sanford NN, Pursley J, Noe B, Yeap BY, Goyal L, Clark JW et al (2019) Protons versus photons for unresectable hepatocellular carcinoma: liver decompensation and overall survival. Int J Radiat Oncol Biol Phys 105:64–72

    CAS  Article  Google Scholar 

  9. 9.

    Kim TH, Park JW, Kim BH, Kim H, Moon SH, Kim SS et al (2019) Does risk-adapted proton beam therapy have a role as a complementary or alternative therapeutic option for hepatocellular carcinoma? Cancers 11:2

    Google Scholar 

  10. 10.

    Cozzi L, Comito T, Fogliata A, Franzese C, Tomatis S, Scorsetti M (2018) Critical appraisal of the potential role of intensity modulated proton therapy in the hypofractionated treatment of advanced hepatocellular carcinoma. PloS One 13:e0201992

    Article  Google Scholar 

  11. 11.

    Chanyavanich V, Das S, Lee W, Lo JY (2011) Knowledge-based IMRT treatment planning for prostate cancer. Med Phys 38:2515–2522

    Article  Google Scholar 

  12. 12.

    Zhu X, Ge Y, Li T, Thongphiew D, Yin FF, Wu Q (2011) A planning quality evaluation tool for prostate adaptive IMRT based on machine learning. Med Phys 38:719–726

    Article  Google Scholar 

  13. 13.

    Good D, Lo J, Lee R, Wu QJ, Yin FF, Das SK (2013) A knowledge-based approach to improving and homogenising intensity modulated radiation therapy planning quality among treatment centers: an example application to prostate cancer planning. Int J Radiat Oncol Biol Phys 87:176–181

    Article  Google Scholar 

  14. 14.

    Moore KL, Brame RS, Low DA, Mutic S (2011) Experience-based quality control of clinical intensity modulated radiotherapy planning. Int J Radiat Oncol Biol Phys 81:545–551

    Article  Google Scholar 

  15. 15.

    Appenzoller LM, Michalski JM, Thorstad WL, Mutic S, Moore KL (2012) Predicting dose-volume histograms for organs-at risk in IMRT planning. Med Phys 39:7446–7461

    Article  Google Scholar 

  16. 16.

    Ge Y, Wu QJ (2019) Knowledge-based planning for intensity-modulated radiation therapy: a review of data-driven approaches. Med Phys 46:2760–2775

    Article  Google Scholar 

  17. 17.

    Fogliata A, Wang PM, Belosi F, Clivio A, Nicolini G, Vanetti E et al (2014) Assessment of a model based optimization engine for volumetric modulated arc therapy for patients with advanced hepatocellular cancer. Radiat Oncol 9:236

    Article  Google Scholar 

  18. 18.

    Yu G, Li Y, Feng Z, Tao C, Yu Z, Li B et al (2018) Knowledge-based IMRT planning for individual liver cancer patients using a novel specific model. Radiat Oncol 13:52

    Article  Google Scholar 

  19. 19.

    Delaney AR, Dahele M, Tol JP, Kuijper IT, Slotman BJ, Verbakel W (2017) Using a knowledge-based planning solution to select patients for proton therapy. Radiother Oncol 124:263–270

    Article  Google Scholar 

  20. 20.

    Delaney AR, Verbakel WF, Lindberg J, Koponen TK, Slotman BJ, Dahele M (2018) Evaluation of an automated proton planning solution. Cureus 10:e3696

    PubMed  PubMed Central  Google Scholar 

  21. 21.

    Vassiliev ON, Wareing TA, McGhee J, Failla G, Salehpour MR, Mourtada F (2010) Validation of a new grid-based Boltzmann equation solver for dose calculation in radiotherapy with photon beams. Phys Med Biol 55:581–598

    Article  Google Scholar 

  22. 22.

    Nocedal J, Wright SJ (2006) Numerical optimisation, 2nd edn.. ISBN 978-0-387-30303‑1

    Google Scholar 

  23. 23.

    Langendijk JA, Boersma LJ, Rasch CRN, van Vulpen M, Reitsma JB, van der Schaaf A et al (2018) Clinical trial strategies to compare protons with photons. Semin Radiat Oncol 28:79–87

    Article  Google Scholar 

  24. 24.

    Langendijk JA, Lambin P, De Ruysscher D, Widder J, Bos M, Verhaij M (2013) Selection of patients for radiotherapy with protons aiming at reduction of side effects: the model-based approach. Radiother Oncol 107:267–273

    CAS  Article  Google Scholar 

  25. 25.

    Rwigema JM, Langendijk JA, van der Laan HP, Lukens JN, Swisher-McClure SD, Lin A (2019) A model-based approach to predict short-term toxicity benefits with proton therapy for oropharyngeal cancer. Int J Radiat Oncol Biol Phys 104:553–562

    Article  Google Scholar 

  26. 26.

    Blanchard P, Wong A, Gunn GB, Garden AS, Mohamed ASR, Rosenthal DI et al (2016) Toward a model-based patient selection strategy for proton therapy: external validation of photon-derived normal tissue complication probability models in a head and neck proton therapy cohort. Radiother Oncol 121:381–386

    Article  Google Scholar 

  27. 27.

    Cheng Q, Roelofs E, Ramaekers BL, Eekers D, van Soest J, Lustberg T et al (2016) Development and evaluation of an online three-level proton vs photon decision support prototype for head and neck cancer—comparison of dose, toxicity and cost-effectiveness. Radiother Oncol 118:281–285

    Article  Google Scholar 

  28. 28.

    Prayongrat A, Kobashi K, Ito YM, Katoh N, Tamura M, Dekura Y et al (2019) The normal tissue complication probability model-based approach considering uncertainties for the selective use of radiation modality in primary liver cancer patients. Radiother Oncol 135:100–106

    Article  Google Scholar 

  29. 29.

    International Commission on Radiation Units & Measurements (2010) Prescribing recording and reporting photon beam intensity modulated radiation therapy (IMRT) (ICRU Report 83)

    Google Scholar 

  30. 30.

    Mizuhata M, Takamatsu S, Shibata S, Bou S, Sato Y, Kawamura M et al (2018) Respiratory-gated proton beam therapy for hepatocellular carcinoma adjacent to the gastrointestinal tract without fiducial markers. Cancers 10(2):58

    Article  Google Scholar 

  31. 31.

    Zhang Y, Huth I, Weber DC, Lomax AJ (2019) Dosimetric uncertainties as a result of temporal resolution in 4D dose calculations for PBS proton therapy. Phys Med Biol 64:125005

    Article  Google Scholar 

  32. 32.

    Zhang Y, Huth I, Weber DC, Lomax AJ (2018) A statistical comparison of motion mitigation performances and robustness of various pencil beam scanned proton systems for liver tumour treatments. Radiother Oncol 128:182–188

    Article  Google Scholar 

  33. 33.

    Bernatowicz K, Zhang Y, Perrin R, Weber DC, Lomax AJ (2017) Advanced treatment planning using direct 4D optimisation for pencil-beam scanned particle therapy. Phys Med Biol 62:6595–6609

    Article  Google Scholar 

  34. 34.

    Zhang Y, Huth I, Wegner M, Weber DC, Lomax AJ (2016) An evaluation of rescanning technique for liver tumour treatments using a commercial PBS proton therapy system. Radiother Oncol 121:281–287

    Article  Google Scholar 

  35. 35.

    Zhang Y, Knopf AC, Weber DC, Lomax AJ (2015) Improving 4D plan quality for PBS-based liver tumour treatments by combining online image guided beam gating with rescanning. Phys Med Biol 60:8141–8159

    CAS  Article  Google Scholar 

  36. 36.

    Bernatowicz K, Lomax AJ, Knopf A (2013) Comparative study of layered and volumetric rescanning for different scanning speeds of proton beam in liver patients. Phys Med Biol 58:7905–7920

    CAS  Article  Google Scholar 

  37. 37.

    Zhang Y, Boye D, Tanner C, Lomax AJ, Knopf A (2012) Respiratory liver motion estimation and its effect on scanned proton beam therapy. Phys Med Biol 57:1779–1795

    Article  Google Scholar 

  38. 38.

    Poulsen PR, Eley J, Langner U, Simone CB, Langen K (2018) Efficient interplay effect mitigation for proton pencil beam scanning by spot-adapted layered repainting evenly spread out over the full breathing cycle. Int J Radiat Oncol Biol Phys 100:226–234

    Article  Google Scholar 

  39. 39.

    Pfeiler T, Ahmad KD, Ayadi M, Bäumer C, Blanck O, Chan M et al (2018) Motion effects in proton treatments of hepatocellular carcinoma-4D robustly optimised pencil beam scanning plans versus double scattering plans. Phys Med Biol 63:235006

    CAS  Article  Google Scholar 

Download references

Author information



Corresponding author

Correspondence to Luca Cozzi.

Ethics declarations

Conflict of interest

L. Cozzi, who is Clinical Research Scientist at Humanitas Cancer Center, acts as Scientific Advisor to Varian Medical Systems. R. Vanderstraeten is senior product manager for proton treatment planning at Varian Medical Systems. A. Fogliata, F.-L. Chang and P.-M. Wang declare that they have no competing interests.

Caption Electronic Supplementary Material


Supplementary materials contains explanatory figure about the principal component analysys; additional figure about the model training; comparative figure for the average DVH for more OARs; the side analysis for small vs large lesions.

Rights and permissions

Reprints and Permissions

About this article

Verify currency and authenticity via CrossMark

Cite this article

Cozzi, L., Vanderstraeten, R., Fogliata, A. et al. The role of a knowledge based dose–volume histogram predictive model in the optimisation of intensity-modulated proton plans for hepatocellular carcinoma patients. Strahlenther Onkol 197, 332–342 (2021).

Download citation


  • Intensity-modulated proton therapy
  • Volumetric modulated arc therapy
  • RapidArc
  • Hepatocellular cancer
  • Machine learning