Outcome-based multiobjective optimization of lymphoma radiation therapy plans

Abstract

At its core, radiation therapy (RT) requires balancing therapeutic effects against risk of adverse events in cancer survivors. The radiation oncologist weighs numerous disease and patient-level factors when considering the expected risk–benefit ratio of combined treatment modalities. As part of this, RT plan optimization software is used to find a clinically acceptable RT plan delivering a prescribed dose to the target volume while respecting pre-defined radiation dose–volume constraints for selected organs at risk. The obvious limitation to the current approach is that it is virtually impossible to ensure the selected treatment plan could not be bettered by an alternative plan providing improved disease control and/or reduced risk of adverse events in this individual. Outcome-based optimization refers to a strategy where all planning objectives are defined by modeled estimates of a specific outcome’s probability. Noting that various adverse events and disease control are generally incommensurable, leads to the concept of a Pareto-optimal plan: a plan where no single objective can be improved without degrading one or more of the remaining objectives. Further benefits of outcome-based multiobjective optimization are that quantitative estimates of risks and benefit are obtained as are the effects of choosing a different trade-off between competing objectives. Furthermore, patient-level risk factors and combined treatment modalities may be integrated directly into plan optimization. Here, we present this approach in the clinical setting of multimodality therapy for malignant lymphoma, a malignancy with marked heterogeneity in biology, target localization, and patient characteristics. We discuss future research priorities including the potential of artificial intelligence.

Introduction

Lymphomas constitute a group of B-cell and NK/T-cell neoplasms with widely differing clinical presentations, manifestations, and prognoses.¹ Treatment consists of systemic therapy (cytotoxic drugs, antibodies [unconjugated or conjugated], targeted small molecules, immune therapies) and/or radiation therapy (RT). Most lymphoma subtypes are highly radiosensitive, and RT remains the single most effective treatment modality with respect to local control. The radiation doses needed for curative treatment of lymphomas are significantly lower than those generally needed for most solid tumors. However, RT, even at these doses, is associated with risks of serious long-term side-effects adding to the long-term effects of the systemic treatments. It is, therefore, of the utmost importance to titrate the intensity of all treatment modalities to achieve an optimal trade-off between risk of serious side-effects and chance of long-term survival. It seems self-evident that the radiation doses to all organs at risk (OARs) should be kept “as low as reasonably achievable”. However, it is not the radiation doses, but the toxicity risks associated with different dose levels, that are the real concern. Additionally, it is not clear what “reasonably achievable” means in an individual case.

Clinical vignette

The complexities in our daily clinical RT planning practice can be illustrated by a typical case story:

A 19-year-old female was diagnosed with Hodgkin lymphoma, nodular sclerosis subtype, clinical Stage IIEA with involvement of both sides of the neck, the right subpectoral region, the mediastinum including the thymus, and the right hilum. She was treated with four cycles of adriamycin, bleomycin, vinblastine, and dacarbazine (ABVD) and achieved a complete metabolic remission after two cycles of ABVD. At the end of chemotherapy, she developed pulmonary infiltrates typical of bleomycin toxicity, had a reduction in diffusing capacity of the lung for carbon monoxide (DLCO) to 66 %, but had no subjective symptoms. She was planned for involved node RT to a total dose of 30.6 Gy in 17 fractions, carried out in deep inspiration breath hold according to the principles of the International Lymphoma Radiation Oncology Group (ILROG) guidelines.² Two coronal slices of the planning CT-scan are shown in Figure 1 demonstrating the clinical target volume (CTV) for treatment planning.

Figure 1.

Clinical target volume on two coronal slices of a planning CT-scan of a 19-year-old female with Hodgkin lymphoma, nodular sclerosis subtype, clinical Stage IIEA with involvement of both sides of the neck, the right subpectoral region, the mediastinum including the thymus, and the right hilum. CTV, clinical target volume: pink; heart: brown; lungs: blue; thyroid: light green.

In such a situation, CTV coverage with the prescribed dose is desirable and existing powerful RT planning tools are capable of creating several distinct conformal plans depending on the dose planner’s list of priorities. The goal is to generate the plan that offers the highest chance of permanent cure and the lowest risk of serious long-term side-effects. Table 1 shows the factors that should be considered in RT planning for this very young patient.

Table 1.

Factors that should be considered in clinical practice for treating a 19-year-old female diagnosed with Hodgkin lymphoma, nodular sclerosis subtype, clinical Stage IIEA

Adverse outcome		Influential factors and associated considerations
Secondary malignancies	Breast cancer	radiation dose to the breasts the prognosis of a secondary breast cancer, should it occur, considering the effect of the patient following the recommended breast cancer screening program
	Lung cancer	radiation dose to the lungs smoking habits (increased risk almost exclusively in smokers) the prognosis of a secondary lung cancer, should it occur
	Thyroid cancer	radiation dose to the thyroid the prognosis of a secondary thyroid cancer, should it occur (they are nearly always well differentiated with an excellent prognosis.)
	Other cancers, e.g. skin cancer, head & neck cancer, connective tissue cancers (sarcomas)	radiation dose to the relevant organs at risk the prognosis of such a secondary cancer, should it occur
Heart disease	Myocardial infarction	mean heart dose or the dose to the individual coronary arteries the prognosis of a myocardial infarction, should it occur
	Valvular heart disease	mean dose to each of the valves the prognosis of valvular disease (primarily aortic and mitral), should it occur
	Arrythmias	heart dose the prognosis, should it occur
Stroke		mean dose to carotid arteries the prognosis of a stroke, should it occur.
Radiation pneumonitis and chronic pulmonary damage		mean lung dose and lung V5a, V20, and V30 risk factors, such as smoking and bleomycin lung toxicity the prognosis (it is very good for pneumonitis but less so for chronic pulmonary damage.)
Hypothyroidism		radiation dose to the thyroid the prognosis of the condition considering the effect of the patient following the recommendation of yearly thyroid testing and thyroid hormone substitution when indicated
Muscular atrophy of the neck with pain and muscle weakness		this is seen many years after mantle field irradiation and dose response data for it are extremely scarce.
Lymphoma recurrence		if some part of the target is deliberately left untreated or undertreated to avoid excessive risks of toxicity from a particular normal tissue the prognostic impact of such an increase in lymphoma recurrence risk

^aVx stands for volume receiving at least X Gy.

Treatment planning complexity

Table 1 is an example of the extensive dose–response information required for treatment planning. Obviously, no one can remember all such information for all relevant outcomes, let alone combine and process this information for each proposed treatment plan. For clinical purposes, simplified constraints for the most important long-term effects are used, without a quantitative risk estimate considering the whole dose spectrum for each long-term effect, and without a quantitative overall estimate considering various long-term effects simultaneously. Treatment planning and plan approval have therefore been based mostly on personal experience and qualitative information influenced by prominent issues, such as secondary breast cancer in the above clinical vignette. Hence, RT planning remains to a large extent subjective and can undoubtedly be improved by:

• Collecting data that are suitable to formulate dose-response relationships for different long-term effects in OARs and enable risk calculations for clinical usage. To this end, dose distributions on large numbers of treatment plans should be analyzed using consistent contours for all the relevant OARs. This is a monumental task for which the use of automation and artificial intelligence (AI) is highly promising.³ AI algorithms are especially useful when the amount of data to analyze is beyond human capacity, or some variables of interest or variable-interplays are not detectable/assessable by a human analyzer.

• Developing mathematical tools for calculating and combining the relevant risk estimates associated with each treatment plan, and ultimately, optimizing plans based on a multiobjective outcome-based risk estimate. At present, we have insufficient data to build high-quality mathematical tools. However, conventional qualitative plan evaluation relies on the same insufficient data, and their usage in a systematic and quantitative way will lead to a consistent and reproducible planning strategy.

The variation in age distribution among the different lymphoma subtypes is wide. Some subtypes, like diffuse large B cell lymphoma, mostly affect the elderly, while some other subtypes, notably Hodgkin lymphoma, have a peak incidence in adolescents and young adults. The histologic subtype is also strongly associated with the anatomic disease distribution, hence some disease locations like the mediastinum are much more common in younger than in older patients. This fact, combined with a very good prognosis with modern treatment, means many of these patients become long-term survivors. For these young patients, the risk of serious long-term side-effects, particularly secondary cancers^4,5 and cardiac diseases,^6–8 is a major concern. However, for older patients, tolerating the high-intensity curative treatment is a major concern. For most lymphoma subtypes, the prognosis is much poorer in older than in younger patients. Nevertheless, older patients tolerate RT rather well, particularly the modern highly conformal and limited-volume RT. In order to generate a personalization paradigm based on patient outcome risks and adjusted by patient preferences, the RT plan and the whole treatment should be tailored to the patient’s lymphoma subtype; the anatomic location and disease extent; the response to chemotherapy, if given; the tolerance of the surrounding OARs; the patient’s age, sex, comorbidities, other risk-factors; and the patient’s own preferences (e.g. Behringer et al⁹). Figure 2 shows such data that should be integrated in RT plan optimization.

Figure 2.

Covariates that need to be included in an outcome-based multiobjective optimization paradigm for radiotherapy planning.

Genomics, in particular germ-line single nucleotide polymorphisms (SNPs), has attracted considerable interest as a potential cause of inter individual variability in early and late toxicity after RT. While the early literature was dominated by false-positive findings,¹⁰ there is now emerging evidence that sequence alterations may affect post-RT adverse events. As genotyping becomes increasingly a part of routine clinical work-up, it is possible that novel SNPs of clinical importance will be discovered in the future. However, except for a few rare genetic disorders, at the time of writing, there are no genotypes that are routinely used in modifying RT prescriptions in the clinic.¹¹ If such genotypes were identified and validated, it would be straightforward to incorporate these directly as covariates in toxicity models for individual patients.¹²

Tailoring treatment to optimize outcomes widens the scope of RT plan optimization. Traditionally in RT practice, a patient’s plan is optimized only with respect to dosimetric aims based on data collected over years from the RT patient population. More recently, however, patient-specific risk-factors and several actionable biomarkers, such as serum and image markers collected before/during the course of treatment, have proven valuable in predicting treatment outcomes. These data are, therefore, potentially important in adapting a plan to the patient’s specific health risks and treatment response.¹³ Accounting for such factors in treatment requires integrating them in RT plan optimization, a process that adds to RT planning’s computational expense.¹⁴ While automation has been a highly desired feature in RT planning,^15,16 the above-mentioned interests have raised this demand to an even higher level.

Current radiotherapy planning

The current RT planning paradigm is schematically illustrated in the left panel of Figure 3. Here, dose–volume constraints for each OAR are defined from population data, and the dose planner and treating physician make a qualitative assessment of the optimal dose delivery considering achievable compromises between target coverage and OARs’ exposure. There are two critical limitations to this approach: (i) it relies solely on population data without incorporating patient-specific risk-factors when assessing toxicity profile, and (ii) target coverage requirements are population-defined (usually consensus-driven) instead of data-driven towards the more appropriate end goal of longest possible overall survival for each patient (under personalized preferences for associated risks).

Figure 3.

A schematic workflow of the current dose-based treatment planning (left) and radiation dose planning using plan libraries (middle) and outcome-based multiobjective optimization (right). We also show in which parts artificial intelligence algorithms can impact the current workflow (grey box on left). AI, artificial intelligence; RT, radiation therapy; TPS, treatment planning system.

For decades, population-based dose-volume constraints, as reported in QUANTEC,¹⁷ offered the most practical approach to quantitative OAR tolerance doses. However, radiation oncologists have increasingly recognized the inadequacy of population-based metrics and tried to adjust RT plans with respect to an individual patient’s health status and risk-factors. Yet, presently, these decisions are rarely made quantitatively. Such a process undertaken subjectively is sensitive to human error and lack of experience. In addition, it is difficult to make a trade-off, and to quantify and compare different solutions, if there are several outcomes to consider (see the Clinical vignette in the Introduction). For lymphoma RT in particular, planning approaches with fixed population-based tissue tolerance doses may provide a suboptimal treatment plan since fixed dose-constraints may well be above what is achievable depending on target localization and prescription dose. While the need of RT planning to go beyond simple dose-constraints by embracing patient-specificity has become evident, clinically feasible ways to do so have not been implemented in routine practice.

There are also practical limitations with the current RT planning workflow: it is time-consuming, repetitive in various steps and largely operator dependent. These limitations have inspired automation of several steps; e.g. contouring an organ or a sub volume of an organ,^18,19 estimating three-dimensional dose distribution,²⁰ producing one plan or a set of possible RT plans for each patient^21,22 or performing a combination of these tasks.²³ As illustrated in Figure 3 (left), advanced computational techniques, including AI, can offer powerful tools for such automation.^16,24 For instance, convolutional neural networks have been used to model and predict voxel-based dose distribution with the ultimate goal of automating RT plan generation.^23,25 However, one obstacle is the lack of an algorithmic definition of what makes a human planner modify a plan in one direction rather than another, particularly in non-trivial cases (the Clinical vignette in the Introduction shows one such case). This obstacle has caused inability to model the human decision-making process.²⁶ Instead, AI agents are mostly designed to use curated data and learn/model patterns.²⁷ Such models, however, tend to have a black box nature presenting interpretability issues to human evaluators.²⁶

Finally, for some lymphoma subtypes, the necessity of RT administration is debated. Lymphoma heterogeneity makes the decision on administering RT and RT dose prescription dependent on both patient and disease specifications. Such challenging decisions cannot rationally be made without an assisted/smart decision-making tool for:

• what combinations of treatments a patient should have^28,29

• whether RT should be included³⁰

• whether ultra-low dose RT should be considered^31–34

• what RT technique should be selected at different stages³⁵

Challenges of radiation therapy for various subtypes of lymphoma

The challenges of lymphoma RT are to some extent different from the challenges of RT for most solid tumors due to the wide variations in (A) prescribed doses and (B) target volumes.

1. Prescribed doses may vary from 4to 50 Gy, a more than 10-fold variation, which is very unusual compared to other tumor histologies:

   1. Localized indolent lymphomas are treated with curative intent to a total dose of 24 Gy.^36,37 However, these lymphomas are exquisitely radiosensitive, and doses as low as 4 Gy will provide durable local control in 2/3 of patients.³⁸ This low dose is used routinely for palliative treatment of disseminated disease,^39–42 but increasingly also in selected patients with localized disease, and only patients without complete remission are offered additional RT to the standard dose of 24 Gy.
   2. Aggressive lymphomas are treated with multiagent systemic therapy with/without consolidation RT. These lymphomas are less radiosensitive compared to the indolent lymphomas, and doses of 30 Gy are used if complete remission is achieved with the systemic treatment.³⁶ If there is still viable lymphoma, doses of 40 Gy or even higher are used.³⁷ Some types of aggressive lymphomas such as extranodal NK/T-cell lymphomas may need radiation doses of 45–50 Gy.^43–46
   3. Classical Hodgkin lymphomas are treated with multiagent systemic therapy with/without consolidation RT. For early stage patients without risk-factors, a radiation dose of only 20 Gy is used,⁴⁷ whereas for early stage patients with risk-factors, doses of 20–30 Gy are used depending on the intensity of the systemic therapy regimen.⁴⁸ For patients with advanced disease showing residual lymphoma after systemic therapy, doses of 36–40 Gy are advised.^2,49 The rare subtype of nodular lymphocyte predominant Hodgkin lymphoma is treated with RT alone in early stage disease, and the dose is usually 30 Gy.²

2. The target volume for lymphoma RT is extremely variable as lymphomas may arise everywhere in the body, both within and outside the lymphatic system. Unlike most solid tumors, where curative RT is often constrained to one particular organ or tissue of origin, any organ may need to be irradiated, and any tissue may be close to the target and, therefore, at risk of acute and long-term side-effects. Moreover, the disease may be localized or widely disseminated, and the target volume needed to be treated may vary from a few to several hundred (or even thousand) cm³.

   1. Some lymphoma subtypes are most often located within the lymphatic system but may be localized in 1–2 lymph node regions or widely disseminated to many regions. RT for lymphomas presenting within the lymphatic system is administered according to the ILROG guidelines for Hodgkin² and for nodal non-Hodgkin lymphomas.³⁷

   2. Some lymphoma subtypes are most often located outside the lymphatic system, where the presenting lesion is extranodal and where the extranodal lesion constitutes the predominant disease bulk.⁵⁰ These represent about 20–25% of all lymphomas, and may arise anywhere outside the lymph nodes, most commonly the gastrointestinal tract, skin, Waldeyer’s ring, central nervous system, salivary glands, and ocular adnexae. RT guidelines for these very variable lymphomas have been developed by ILROG.^44,46,51

This extraordinary variation in target and dose prescription for lymphomas generates an almost infinite number of possible combinations, each associated with risks of several OAR complications. The common, rather crude use of dose-constraints for treatment planning is clearly inadequate; since for some patients, it is impossible to keep within the constraints, and, compromises must be made, whereas for others with less aggressive disease, much lower doses may be effective. This variation disqualifies the standard dose constraints as indicators of plan quality.

Plan-library and Paretofront-based radiotherapy planning

In recent years, an RT planning technique based on the direct use of “good” plans from a historical library of previously treated patients, informing the computer on achievable planning goals, has been investigated (Figure 3, central panel). The aim is both to save human planners from tedious repetitive work and to reduce the risk of suboptimal planning due to lack of experience of an individual dose planner. For instance, the concept may suggest achievable OAR exposure levels based on the individual patient anatomy,⁵² and subsequently suggest optimization objectives and weights to reach corresponding “good” dose distributions in the patient. Hence, these procedures may have the potential to reduce the risk of suboptimal dose distributions, e.g. dose distributions that fulfill population-based constraints but could still be improved. Also, these methods may be expanded to include the entire planning stage (selection of energies, field angles, etc.) on top of the inverse planning process itself in order to reach a high level of automation.⁵³ The approach has been named “knowledge-based planning”, but such wording is too unspecific in the current context.

Several groups have reported encouraging results with the plan-library approach in a number of disease sites, such as head and neck,²² prostate,⁵⁴ uveal melanomas⁵⁵ and female supradiaphragmatic Hodgkin lymphoma.⁵⁶ The technique has also been commercially implemented and further gained AI-assisted automation in some steps in the research studies.⁵⁷ For example, in prostate RT, Boutilier et al⁵⁸ and Ma et al⁵⁴ used AI to model and predict optimum priority weights for dose–volume constraints, and to predict dose–volume histograms, respectively. For lymphoma RT, in many cases, it is possible to share algorithms derived from plan libraries between institutions potentially leading to improved plan quality globally. However, conclusive data are currently lacking, both at the institutional and societal level. It would be desirable if algorithms were published in a way that can be applied in other institutions and benchmarked against state-of-the-art manual planning processes in leading institutions.

Regardless of prior planning experience, prioritizing conflicting optimization objectives is a challenge (a primarily numerical challenge) to finding solutions that are “as good as possible”. Such challenges arise in multiobjective optimization in many applied sciences, including engineering and economics, where optimal decisions need to be taken in the presence of trade-offs between two or more conflicting objectives that are incommensurable. A nontrivial multiobjective optimization problem has no single solution that simultaneously optimizes each objective. Instead, there is a generally infinite set of solutions that are Pareto-optimal, i.e. solutions where none of the objectives can be improved without deteriorating one or more of the other objectives. A simple two-objective case would be the optimization of target dose-coverage versus dose to a single OAR. In non-trivial cases, the two objectives conflict, in the sense that target coverage can be improved but only at the expense of increasing the dose – using some biologically relevant dose–volume metric – to the OAR or vice versa. A clinical decision is required for prioritizing the two objectives. In practice, the space of possible solutions is infinite, but some of these plans are objectively worse than others. Assume that a plan, Π, achieves tumor dose-coverage Dx with OAR exposure, d. Then, all plans with the same dose-coverage Dx, but higher OAR dose than d are objectively suboptimal, they are dominated by Π. Likewise, Π dominates any plan with the same or worse d that achieves the same or worse Dx. The Pareto-front defines all Pareto-optimal solutions. In clinical decision-making, the solution-space of interest should, therefore, be restricted to solutions on the Pareto-front (see an illustrative example in Figure 4). In the general clinical scenario, the Pareto-front will be n-dimensional with n > 2, i.e. with at least one target-related criterion and typically >1 normal-tissue related criteria. Note, however, that no reference is made to dose-constraints. In other words, if a plan is on the Pareto-front, it will not have more than necessary OAR exposure for a given target coverage chosen by the treating physician.⁵⁹ However, not all the plans on the Pareto-front meet the pre-specified clinical constraints.

Figure 4.

A simple Pareto-front for two objectives: (1) a tumor/target objective and (2) an OAR objective. The darkened areas highlight the space that does not meet either the clinical constraint for tumor coverage or for dose to an OAR. The Pareto-front defines the deliverable radiation dose plans that cannot be improved on one objective without deteriorating the other. OAR, organ at risk.

Pareto-front planning is included in some clinical plan optimization software tools using a principal component analysis method, first introduced by Craft et al,⁶⁰ to reduce the dimensionality of the Pareto-front. Breedveld et al⁶¹ showed that for an already chosen set of priorities for optimization objectives, the choice of the best plan on the Pareto-front can be performed automatically.

Deciding between Pareto-optimal plans in lymphoma RT is complicated and requires trade-offs made by the patient, the physician or both in shared decision-making (See the clinical vignette in the Introduction demonstrating an example of such complexity).

Outcome-based multiobjective optimization for radiotherapy planning

Moore very appropriately stated RT planning challenges: “Treatment planning is hard because the line between a good and bad plan must be uniquely negotiated for each patient, and most treatment planning processes do not sufficiently constrain the possible solution space.”¹⁵

Figure 5 demonstrates a limitation of defining the Pareto-front in the space of physical dose delivered. Patients with different disease characteristics may be subject to substantially different probabilities of treatment failure or toxicity risk for the same given dose; therefore, a Pareto-optimal plan in dose space may not be clinically acceptable in the projected outcome for the patient.

Figure 5.

Translation of a Pareto-front from physical space to risk space will depend on the patient’s risk factors. Here is an example where a Pareto-optimal dose delivers a good risk profile in a patient without cardiac risk factors (point 1). However, the addition of a risk factor (pre-existing cardiac conditions) will move the Pareto-front to the right by an amount proportional to the relative risk (shown as α RR in the right panel) associated with the risk factor which may render the plan unacceptable in risk of toxicity (point 2). Typically, a compromise will be made but a qualitative compromise may render a non-Pareto-optimal plan (point 3) which can be improved in disease control without changing the risk of toxicity (point 4).

Here, we define RT planning with outcome-based multiobjective optimization (RT-OMO) with reference to modeling of short- and long-term treatment outcomes (or outcome surrogates) and using these models as an integrated part of the plan optimization process. The utility of tumor control probability (TCP) and normal tissue complication probability (NTCP) models for this purpose allows personalization at the model stage, i.e. the incorporation of patient-related risk-factors in the models. Such probabilistic models can be generated analytically via statistical analysis or by an AI-based algorithm. While, in principle, RT-OMO can be used within library-based planning and Pareto-front optimization frameworks (see the right panel of Figure 3), it has so far only been rudimentarily explored in research papers. Brodin et al presented a precursor of RT-OMO by employing existing NTCP models and allowing clinicians to observe the effect of each toxicity individually and choose a “good” plan.⁶²

Figure 6 gives a brief illustration of the modeling process for a patient with a risk-factor. The effect of some factors such as age at diagnosis, a factor that has been routinely registered for all patients and continuously found significant in treatment response, can be readily modeled using large existing data bases. Several other risk-factors, however, are more sporadically recorded in clinical registries and need systematic, prospective data collection before their significance in a patient’s treatment response can be established. Individualization is relevant for both NTCP and TCP estimates. Examples are the impact of co-morbidities on the risk of cardiac mortality,^63,64 smoking on the risk of secondary cancers^65–67 and age on the risk of lung toxicity.⁶⁸

Figure 6.

An illustration of modeling workflow for quantification of the effect of a risk-factor on RT outcome is shown. The orange box highlights where, at present, this workflow’s largest weakness resides. RT, radiation therapy.

There are also patient-specific factors with partly known effects as response modifiers. In the current RT practice, these factors are monitored via surrogate markers and can trigger interventional adjustments as response-based adaptation. Response-based RT planning is again interesting but far from clinical implementation in most solid tumors. Lymphomas are special by having a large evidence base for using response to chemotherapy to modify the subsequent treatment in several situations.^13,69,70 Lim et al⁷¹ suggested data collection to investigate the use of interim functional imaging in guiding the choice of RT in addition to chemotherapy for advanced-stage Hodgkin lymphoma. Similarly, GAZAI trial investigates minimizing the risk of higher recurrence rate after low dose RT by additional salvage radiation up to 40 Gy for patients who have less than a complete response.⁷² Another potential adaptation comes with the use of blood biomarkers in outcome risk analysis which can, subsequently, be applied in RT-OMO.^73,74

Several studies have either created patient-specific models for a particular treatment-related toxicity or modeled the effect of a single patient-specific factor on survival.^6,74,75 While informative, such modeling efforts cannot in themselves guide RT planning toward a “good” plan. RT-OMO aims at developing a full mathematical representation of patient-specific risks of post-RT outcomes. Rechner et al introduced a preliminary version of one such planning approach by modeling a proxy for all-cause mortality in mediastinal Hodgkin lymphoma and fully optimizing RT plans by minimizing this end point.¹⁴ Similar to the approaches taken by optimizing for life-years-lost due to multiple adverse outcomes⁷⁶ or failure to control the malignant disease,⁷⁷ Rechner et al.’s study combined multiple endpoints into one scalar – a step that allowed the practical implementation of the proposed RT-OMO.

Implementation of RT-OMO also involves quantification of the associated uncertainties.⁷⁸ Since the outcome risk models provide the required link to move from dose to risk profiles, the uncertainties in these models can cause significant variation in planning results. Additionally, the uncertainty analysis can inform which areas need further data collection. Figure 7 summarizes a critical point in outcome-driven optimization presented by Rechner et al¹⁴. The tumor control probability was modeled as the probability of no relapse where generalized equivalent uniform dose (gEUD) was used to present dose to CTV: gEUD = $(\frac{1}{n}\sum _{i=1}^n{d}_i^{\alpha }{)}^{\alpha }$with n being the number of target voxels, d, the voxel dose, and $\alpha$, the gEUD parameter. Figure 7 gives an example of the sensitivity of RT-OMO to the parameters used in the models by demonstrating how the choice of $\alpha$ could drastically change the outcome-optimized plan and its associated dose distribution.

Figure 7.

An example, adapted from in Figure S6 Rechner et al¹⁴’s study, showing the variation in dose distribution in one sagittal and two axial views when different model parameters are used in treatment plan optimization. Left: tumor control model using gEUD α = 1 yields a plan in which mean target dose matters and not dose to each target voxel. Right: tumor control model using gEUD α=-22 yields a plan in which dose to each target voxel matters. gEUD, generalized equivalent uniform dose.

As highlighted in Figure 6, the major roadblock for clinical implementation of a workflow that performs RT planning with outcome-data-integration resides in the middle word: data. There is a need for purposeful data collection that adequately informs outcome risk models. While such data remain unavailable, data-driven techniques are making the best use of the available data regardless of incompleteness. As these techniques get more popular, more informed data collection efforts will emerge to supply the required models.

Discussion

Automation and personalization of lymphoma RT planning is challenging. Lymphoma heterogeneity in terms of volume, shape and initial location allows a large set of treatment strategies and a wide range of RT dose prescriptions. Increased knowledge of dose–effect relationships will lead to more organs to delineate, a task which benefits from AI-assisted techniques to manage the workload.^18,79,80 AI is also promising in mining historical data of exposure and outcome to further expand our knowledge of risk of toxicity or disease control by increasing the number of patients that can be analyzed.^3,81 Finding the optimal treatment in each single patient is highly non-trivial (see the Clinical vignette in the Introduction as an example) and would benefit from all aspects of advanced planning.

Irrespective of disease site, AI-generated results are dependent on the database mined,⁸² hence, these results need final approval by a clinician. Wang et al⁸³ reviewed the studies showing diagnostic and prognostic value of some image-based features for lymphomas and emphasized the suboptimal quality of the studies and the required caution in interpreting their results.

While personalizing the RT planning process is not a new idea, with recent technological and computational advances, it might finally be turning the corner from a theoretical research endeavor to a standardizable clinical workflow for some disease sites. In lymphomas, the achievements have been limited due to the wide spectrum of the diseases and the treatment strategies (See Table 2 for some recent relevant work).

Table 2.

Studies on personalizing RT planning for lymphomas, published since 2018

Institution/First author/Publication year	Lymphoma subtype	RT planning personalization technique
University of Copenhagen (Denmark)/Specht/201849	Hodgkin lymphoma	Review on outcome-risk-based personalization
University of Southampton (UK)/Lim/201871	Hodgkin lymphoma -Advanced stage	Review on risk and response adaptation in RT planning
Cancer campus Grand-Paris (France)/Boros/201835	Hodgkin lymphoma	Personalized choice of RT modality based on risk-factors
University of Torino (Italy)/Levis/201984	Hodgkin lymphoma	Heart substructure doses were used for personalization. The Statistical model introduced by van Nimwegen et al6 was used to estimate the dose response.
Yonsei University College of Medicine (Korea)/Lee/201985	Orbital mucosa-associated lymphoid tissue lymphoma	Studying relapse patterns in low dose RT
MD Anderson Cancer Center (TX)/Milgrom/201974	Mediastinal Hodgkin Lymphoma	Use of PET radiomic features to model relapse probability
Davidoff Cancer Center (Israel)/Siegal/201986	Central nervous system lymphoma	Use of age as a personalization parameter
University of Copenhagen (Denmark) & University of Maryland (MD)/Rechner & Modiri/202014	Hodgkin lymphoma	Outcome-risk-based personalization

In order to benefit from the emerging computational advances for lymphoma RT, cause–effect relationships should be identified and quantified. To that end, the gaps in analyzable data, over which an effective next generation treatment planning approach can be built, need to be filled. The collected data should determine whether it is reasonable to compromise TCP for longer survival and how much TCP compromise is safe in various lymphoma subtypes and in different patients. “Steering between undertreatment, with the risk of avoidable recurrences, and overtreatment, with the risk of unnecessary toxicity, remains complex because control of the lymphoma and the probability of survival are no longer closely linked.”⁷¹ Quantitative assessment of the risks of a multitude of post-RT outcomes should be a key goal, but also a reachable goal, given the advances of state-of-the-art RT.

In conclusion, we are convinced that current qualitative compromises will not result in fully optimal RT plans for the individual lymphoma patient. RT-OMO, as outlined here, will yield Pareto-optimal plans, but it is a research question whether the achievable improvements are trivial (as seen in breast cancer cases studied by Stick et al⁸⁷) or clinically meaningful. Based on the published pilot case studies in Hodgkin lymphoma, we hypothesize, however, that there will be a subset of lymphoma patients where target and patient profile are complex and where personalized, outcome-based multiobjective optimization will provide tangible benefit to the patients.¹⁴ Full realization of the potential of the proposed strategy will require improvements in the quantitative planning techniques. AI and big data analytics are likely to play an important role in facilitating these types of studies.