The aim of radiotherapy treatment planning is to define a treatment that provides adequate tumor volume irradiation with the highest expected therapeutic ratio. To this purpose, doses in organs at risk (OARs) are minimized based on known risks for radiation-induced toxicity (1). Technical developments in external beam radiotherapy (EBRT), e.g., replacement of 3D-conformal radiotherapy (3DCRT) by intensity modulated radiation therapy (IMRT) and volumetric modulated arc therapy (VMAT) (2–4), and improvements in image guidance (5–7), have significantly improved treatment outcome and/or reduced radiation induced side effects in a variety of treatment sites. Recently, developments in automation of treatment planning have further enhanced opportunities for generation of high quality treatment plans (8–10).

Ideally, toxicity risks to be used in planning are modeled with normal tissue complication probabilities (NTCPs). There is an active field of research developing these predictive models (1, 11–15). More and more, published NTCP models include non-dosimetric parameters that modulate the radiation-induced toxicity risk (16). For example, Cozzarini et al. (14) used multivariate logistic regression to include both dosimetric parameters, extracted from the clinical plans, and patient characteristics (e.g., smoking status, age, application and duration of hormonal therapy) in the toxicity prediction models. Pre-selection of a relevant predictor subset was performed using univariate logistic regression. A similar approach was performed in previous work by Sharfo et al. (17) who developed a multivariate logistic regression model predicting radiation induced gastro intestinal (GI) toxicity.

Current practice in radiation therapy treatment planning is based on treatment site specific clinical protocols, containing hard constraints, and planning aims. Evidence based medicine recommends the definition of clinical protocols, based on findings in prospective clinical trials and dose escalation studies (18). Generally, the planning protocol is used by a planner to generate for each patient a single treatment plan that may or may not be adjusted after discussion with the treating physician prior to final approval. There is no systematic exploration of patient-specific trade-offs between the various planning aims by generation of a set of treatment plans for each patient with (slightly) different trade-offs.

We hypothesized that generation of a limited set of well-designed treatment plans per patient, instead of a single plan, can help to better identify plans with optimal patient-specific trade-offs. For example, for some patients with specific anatomies, a slight decrease in coverage might result in a relatively large NTCP gain. For patients with non-dosimetric conditions that result in a significantly enhanced predicted NTCP, a lower PTV coverage or a somewhat enhanced NTCP for a different side-effect might be accepted to counter-act the enhancement. We also hypothesized that automated planning can be used to effectively generate the required treatment plans.

In this paper we have investigated these hypotheses for treatment of prostate cancer. An automated planning algorithm was used to generate for each patient a set of plans to explore the trade-off between the dose coverage of the large planning target volume to be irradiated with reduced dose (PTV_Low) and the predicted NTCP for grade ≥2 GI toxicity for otherwise similar dose distributions. In particular, measures were taken to maintain clinical target volume (CTV) coverage at 100% and to keep the coverage of the (smaller) PTV_High at the requested ≥99% level. Deterioration of bladder dose was also to be avoided. We also investigated to what extent reduction in PTV_Low coverage could compensate for significantly enhanced toxicity risks caused by diabetes.

Figures 1, 2 show NTCP reductions for the 50 study patients as a function of the loss in PTV_Low dose coverage. Figure 1 is valid in case of diabetes, while for Figure 2 we assumed that there was no diabetes. As explained in the M&M section, reductions in PTV_Low coverage in the six alternative plans for each patient were obtained with (convex) LTCP cost functions. Convexity avoids getting trapped in local minima, but with the LTCP cost function, obtained PTV_Low coverage values vary somewhat between patients. For generation of Figures 1, 2, NTCPs for the defined coverage reductions were for each patient obtained by piecewise linear interpolations between the generated plans. The different colors show the impact of incremental underdosage steps of 1% in PTV_Low on obtained NTCP. For some patients (e.g., patient 13), reducing the coverage to as low as 90% was not possible, possibly due to not sufficiently large PRV margins or conflicting constraints on the PTV_High and the CTVs dose requirements. For patient 50, accepting lower PTV_Low coverage did not result in any NTCP reduction because of lack in overlap between PTV_Low and rectum (see also Figure 6).

Following Equation 1, NTCP values were indeed higher in case patients had diabetes (compare upper panels of Figures 1, 2). On the other hand, NTCP reductions were also larger in case of diabetes. For a PTV_Low coverage of 95%, average NTCP reductions of 4.3% (0.3–8.0%) and 2.9% (0.2–5.5%) were obtained with or without diabetes, respectively. For 90% coverage, the obtained NTCP reductions increased to 8.3% (0.3–14.4%) and 5.6% (2.0–10.1%), respectively. Both with and without diabetes, there was an overall trend toward enhanced NTCP reductions for patients with the highest clinical NTCPs (lower panels Figures 1, 2). On the other hand, large inter-patient variations were observed. For example, patients 1 and 3 had similar clinical NTCPs, but a large difference in achievable NTCP reductions. Moreover, similar NTCP reductions were observed for different costs in PTV_Low coverage. For example, patients 12 and 14 have similar NTCP reductions of ~10% accepting 94% or 91% PTV_Low coverage instead of 99% (Figure 1). Observed maximum NTCP reductions ranged from > 14% (patient 1) to <1% for patient 50, depending on differences in anatomy (see Figure 6).

Figures 3, 4 show the differences between clinical and alternative plans on a per patient base and in population DVHs, respectively. They demonstrate that the enforced PTV_Low coverage reductions mainly had an impact on rectum sparing while having a clinically insignificant dosimetric impact on PTV_High, CTV_High, CTV_Low, bladder, anus and hips, as intended (section Generated Treatment Plans). Figure 5 shows for an example patient highly similar dose distributions, except for the region of overlap between rectum and PTV_Low.

In Figure 6 we investigated the extent of feasible NTCP reduction as a function of overlap between rectum and PTV_Low. Although, there is an overall trend toward more reduction with larger overlap, there are inter-patient variations with R² equal to 0.6 and 0.7, for 95 and 90% PTV_Low coverage, respectively.

For the 50 patients in this study, presence of diabetes resulted in an average increase in clinical NTCP from 15.4 ± 3.0% (1SD) to 24.9 ± 4.5% (1SD) (compare also the upper panels of Figures 1, 2). Figure 7 explores opportunities for mitigation of enhanced toxicity risk due to diabetes by reducing required PTV_Low coverage. Clearly, depending on the allowed coverage reduction and the patient anatomy, NTCP enhancements due to diabetes could be largely compensated. For some patients, (e.g., 1, 6, and 8) the impact of diabetes could be completely canceled when using a coverage of 90–91%. Other patients (e.g., 3 and 20) demonstrate quite large residual differences in NTCP with and without diabetes, for reduced PTV_Low coverages.

In this study, we have used prostate cancer radiotherapy as a model for development of a clinically feasible workflow for application of automated planning for assessment of patient-specific trade-offs between treatment goals. All plans were generated fully automatically, i.e., without any manual fine-tuning. With carefully designed, patient-independent variations in the autoplanning configuration (i.e., wish-list), the PTV_Low coverage could be varied in a controlled way in the range 99–90% to reduce the predicted NTCP, without significant further changes in the dose distributions. In particular, CTV coverage remained 100%, PTV_High coverage was kept at ≥99%, and bladder dose did also not significantly change. For each patient, the obtained bladder D_Mean in the clinical plan (PTV_Low coverage ≥99%) was used as constraint in the generation of the six other plans with reduced PTV_Low coverage. It was demonstrated that large, but highly patient-specific NTCP reductions could be obtained. For a PTV_Low coverage of 90%, observed NTCP reductions ranged from 14.4 to 0.1%, compared to 99% coverage, depending on the patient anatomy. Reductions in required PTV_Low coverage could to a large extent make up for diabetes as a co-morbidity, again depending on patient anatomy. To the best of our knowledge, this is the first study that proposes the use of automated planning for patient-specific exploration of opportunities for dosimetric compensation of non-dosimetric toxicity risk factors.

Automated treatment plan generation required about 1–2 h per treatment plan. No manual interaction was required at any step of the procedure. Therefore, multiple plans could be run, sequentially or in parallel, over the night. Generation of a wish-list generally takes several weeks. This is a one-time effort and should be seen as an upfront time-investment, which saves a lot of manual planning time at a later stage. Specifically for this project, the wish-list was already developed in a previous study (17).

Observed NTCP reductions correlated to some extent with the volume of rectum overlapping with PTV_Low (Figure 6, R² = 0.6–07). Once a correlation model is built based on the plans generated with the proposed method, the regression lines might be of use as a tool for selection of the PTV_Low coverage region of interest, or for selection of patients. That is, the proposed method could be applied only to patients and/or to PTV_Low levels that show to be more promising in NTCP reduction. However, even in the relatively easy treatment site of prostate cancer, a not too strong correlation was found. Different parameters may be investigated, but for more challenging treatment sites finding predictors for NTCP reduction may be even more complex. The presented method, on the other hand, only requires computation time once the procedure is defined.

In the wish-lists applied in this study, concentric shells at distances of 5, 15, 25, and 50 mm from the PTV edge were used to control plan conformality (Table 1). The limit and goal values were the same for all patients and all plans. Initially, we did however try to get further NTCP reductions by loosening conformality goal values. This was not successful; conformality worsened but NTCPs remained practically unchanged.

Equation 1 was used for NTCP prediction in this study, as our patients were treated in the context of the HYPRO trial, and Equation 1 was derived for these patients. Important to note is that various alternative predictive models exist (11, 15), which could possibly have resulted in different conclusions, or could have resulted in different approaches for lowering NTCPs. Direct use of Equation 1 in this study was limited to plan evaluations, i.e., Equation 1 was not used in the wish-list for plan generations (see Table 1). For planning, we generally prefer to use convex cost functions to avoid getting trapped in local minima, and the NTCP expression in Equation 1 is not convex. Alternatively, the (convex) rectum gEUD (7.7), as used in Equation 1, was directly applied as an objective function (priority 4 in Table 1).

The proposed method to explore trade-offs in planning goals has some similarities with the well-known Pareto navigation, using a graphical user interface with sliders to find a clinically favorable plan (8, 26–31). Also in that method, multiple plans are automatically generated for manual plan selection. There are, however, important differences. The most important difference is that for each patient, we first generate a high-quality, Pareto optimal plan (“wish-point”) with clinically most desired PTV_Low coverage (≥99%). For each patient, this plan is then used as anchor point for patient-specific generation of the plans with slightly reduced PTV_Low coverage, using the bladder dose obtained in the wish-point plan as constraint. In the proposed workflow, only plans are generated that are useful for the desired analyses. In conventional generation of plans for Pareto navigation, there is no knowledge of the “wish-point,” and generation of plans is less focused. Due to our highly focused plan generation, only few plans are needed for the analyses. In this study we used seven plans per patient. This number was not optimized in terms of finding the minimum number of required plans. The aim was to include for all patients, the full range of PTV_Low coverages from 99 to 90%. If a clinical protocol has more precise directions for reductions in PTV_Low coverage, for sure even fewer plans need to be generated.

PTV margins are generally used to minimize risk in CTV miss. In this paper, we kept all margins unchanged, but allowed doses in the overlap area of PTV_Low with rectum to get lower than in the clinical plan. Coverages in PTV_High and CTV were always maintained. With this approach, the risk of CTV miss was minimized, but still (at least potentially) enhanced compared to regular clinical planning. Therefore, clinical introduction of this type of workflow is not trivial. Extensive computer simulations could be performed to assess the true risks, taking into account the clinically applied image-guided approach. Clinical introduction could well be performed in a formal study. Anyway, it seems that patient selection could be important, with patients with a high clinical NTCP (e.g., related to an unfavorable anatomy or diabetes) and a large potential for NTCP reductions, as best candidates. It is important to realize that we used in this study our clinically required PTV coverage level of 99%. In many studies, coverages of 95% were reported (32).

We have investigated trade-offs between PTV coverage and GI NTCP for prostate cancer but believe that the proposed methodology could also be applied for other tumor sites. The system could also be used to explore patient-specific trade-offs between various toxicities for fixed PTV coverage. Focusing on balances between toxicities instead of toxicity vs. PTV coverage could ease clinical implementation. The developed workflow could potentially also be used in shared decision making studies.

A novel, clinically feasible workflow has been proposed for the use of automated planning to systematically explore patient-specific trade-offs between various treatment aims. For prostate cancer, the patient-specific balance between PTV coverage and predicted GI toxicity risk was explored. Opportunities for compensating significantly enhanced predicted toxicity risk related to diabetes by reducing the PTV coverage were investigated as well. Large variations in potential benefit were observed in the fifty study patients. The proposed system could play an important role in further high-precision personalization of patient care.

The datasets generated for this study are available on request to the corresponding author.

Ethical review and approval was not required for the study on human participants in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required for this study in accordance with the national legislation and the institutional requirements.

RB: conducting main research and writing manuscript. LR and BH: supervision of research and writing manuscript. AS: brainstorming about research and writing manuscript. SB: developing code and writing manuscript. WH: data collection and writing manuscript. LI: PI of Data collection and writing manuscript. All authors contributed to the article and approved the submitted version.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.