A primer on radiobiological modeling for treatment plan optimization

Modern radiotherapy is built on a paradox: the quantity we calculate most precisely is absorbed dose, yet the quantity that determines cure or complication is cellular fate. Radiobiological modeling bridges that gap by converting physical dose distributions into predicted biological outcomes, and it has become the engine behind a generation of biologically guided treatment plan optimization strategies.

For physicists at the treatment planning console, the question is no longer whether to consider biology, but how deeply to embed it. A plan that minimises monitor units or prescription variance is not always the plan that maximises tumour control or spares late-responding organs. Models such as the linear-quadratic formalism and TCP/NTCP frameworks offer a path toward plans that are both mathematically tractable and clinically meaningful.

In Australia, where radiotherapy capacity stretches from the busy metropolitan centres in Sydney and Melbourne to regional hubs in Newcastle and Townsville, these methods carry particular weight. ACPSEM-certified physicists balance tight throughput demands against the desire to personalise plans, and biological objectives can offer a structured way to do both. The framework also resonates with safety oversight from ARPANSA, which encourages evidence-based planning strategies beyond purely geometric checks.

This article walks through the foundations of radiobiological modeling, the practical decisions that go into selecting a model, and the steps required to translate those models into the cost functions that drive an optimiser. It is written for clinical and research physicists preparing to engage with biological optimisation during the Banff scientific program or back at their own consoles.

The biological rationale beyond physical dose

Dose-volume histograms and conformity indices say little about what dose-volume patterns actually mean for a patient. A uniform 70 Gy to a prostate yields a very different biological outcome than a heterogeneous 70 Gy with a 90 Gy hotspot, even when both plans share the same D95. Radiobiological modeling translates those patterns into estimates of cell kill, complication probability, or complication severity.

The core idea is straightforward: every voxel carries information beyond the dose it received, including the tissue type, fractionation schedule, repair capacity, and repopulation kinetics that determine its fate. Once a model is parameterised for those features, the optimizer gains a richer objective than mean dose to a structure.

In Australian practice this rationale aligns with the priorities of the Australian Clinical Dosimetry Service, which audits plan quality across participating centres. Biological metrics can complement the geometric audit framework by highlighting plans that look acceptable on paper but may carry hidden risks for late toxicity.

Survival curves and the LQ model in practice

The linear-quadratic model remains the workhorse of clinical radiobiology. Survival fraction is described by two parameters, alpha and beta, that together capture low-dose radiosensitivity and the curvature of the survival curve. Their ratio, the alpha/beta ratio, is the single number clinicians most often invoke when justifying hypofractionation.

Choosing alpha/beta values is the first modelling decision any physics team makes. Literature values for prostate (around 1.5 Gy) and for many late-responding tissues (3 Gy) differ sharply from values assumed for squamous carcinomas and most tumours (around 10 Gy). The model also breaks down at very high doses per fraction, which has spurred work on modifications such as the universal survival curve and the generalised LQ approach.

Practical users in Australian centres will recognise the same trade-off that comes up in stereotactic lung or liver work at Royal North Shore or the Peter MacCallum: the simple LQ form is easy to implement and fast to compute, but the user must be alert to where its assumptions fail. Newer datasets are refining alpha/beta estimates for many tumour sites, and the radiobiological optimisation community is incorporating those updates routinely.

TCP and NTCP as clinical endpoints

Tumour control probability and normal tissue complication probability functions convert dose distributions into clinical endpoints. TCP models, often built on Poisson statistics of surviving clonogens, predict the probability of eradicating a tumour. NTCP models, including the Lyman-Kutcher-Burman framework and the more recent EUD-based variants, predict the likelihood and severity of complications for organs at risk.

These endpoints are useful because they can be combined into a single biological objective, often expressed as a function such as TCP minus a weighted sum of NTCP values. They also make plan comparisons more intuitive for multidisciplinary teams, since the output speaks the language of patient outcomes rather than of dose-volume constraints.

The parameters of these functions come from historical datasets, often with limited follow-up and heterogeneous endpoints. A physicist building a biologically optimised plan in Brisbane or Perth needs to understand which population informed the model, how endpoints were graded, and whether the model's assumptions still hold for modern techniques such as IMRT or proton therapy.

Translating models into optimization objectives

The leap from radiobiological models to plan optimization is largely a matter of mathematical translation. A cost function built from TCP and NTCP terms can replace the conventional weighted sum of squared dose deviations. Constraints can be written as maximum allowable NTCP or minimum required TCP, and penalty weights can be tied to the clinical importance of each endpoint rather than to dose-volume heuristics.

Some optimizers implement this directly, allowing the user to set biological objectives alongside conventional ones. Others require an intermediate step, where a TCP/NTCP map is calculated from a candidate dose distribution and then fed back into a constrained optimisation loop. Either way, the workflow demands careful documentation, since the assumptions of the model must travel with the plan through quality assurance.

A common Australian pitfall is to adopt biological objectives without revisiting institutional margins. A plan optimised for high TCP and low NTCP can still produce marginal coverage if the model parameters are not aligned with the planning target volume definition used in the clinic. Tight integration between the biology team and the planning team is essential when a centre runs multiple tumour site protocols in parallel.

Validating models against clinical reality

No radiobiological model is self-validating. Predicted TCP and NTCP values must be checked against observed outcomes through registries, prospective audits, and in-silico trials. In Australia, participation in trans-Tasman outcome registries and in initiatives linked to ACPSEM provides an opportunity to ground model-based optimization in real patient data.

Uncertainty is part of the validation exercise itself. Parameter uncertainty, inter-patient variability, and the stochastic nature of radiotherapy outcomes all contribute to wide confidence intervals around any model prediction. A robust optimisation framework can hedge against this uncertainty by encouraging plans whose predicted benefit is relatively insensitive to plausible parameter shifts.

This kind of robustness analysis fits naturally into existing quality assurance workflows. A planner in Adelaide who runs extra scenarios with shifted model parameters can quickly identify whether a high-TCP plan is genuinely advantageous or an artefact of a particular parameter set. Stress-testing biological objectives in this way is one of the most valuable practices a clinic can adopt.

Emerging tools and the path forward

The current generation of radiobiological models is already being reshaped by data-driven methods. Machine-learning approaches can learn TCP or NTCP functions directly from large outcome datasets, sometimes outperforming classical parameterised forms. They also allow integration of features beyond dose, such as imaging biomarkers, genetic profiles, and blood-based indicators of radiosensitivity.

The classical mechanistic models remain useful for their interpretability and for the insights they offer into underlying biology. The most productive research environments treat the two approaches as complementary rather than competing. A voxel-level dose-driven model can be paired with a patient-level learned predictor to capture effects at different scales.

Vendors are responding. Treatment planning systems increasingly expose biological objectives as first-class features, and research platforms allow physicists to plug in custom models. For practitioners, the practical task is to stay close to the literature, participate in working groups coordinated through ACPSEM, and bring biological thinking into everyday planning.

Practical considerations for choosing a model

• Tissue type: late-responding versus early-responding tissues require different model assumptions
• Fractionation schedule: very high or very low doses per fraction may require modified forms of the LQ model
• Endpoint definition: severity grading and follow-up length shape which NTCP function is appropriate
• Computational cost: voxel-by-voxel biological objectives can slow optimization, particularly for arc therapies
• Validation status: prefer models with documented performance in populations similar to the clinic's own

Workflow and data requirements for biological optimization

• Robust outcome data with consistent grading of complications
• Parameter sets that reflect the patient population served, including age and comorbidity profiles
• Integration with the treatment planning system through scripting or native biological modules
• Documentation that captures model version, parameters, and assumptions for every plan
• Regular audits that compare predicted TCP/NTCP against observed outcomes
• Training for dosimetrists and physicists on interpreting biological objective values

Physicists heading to Banff can deepen their grasp of these ideas by reviewing the keynote speaker highlights from the scientific program. Those arranging their stay should consult the latest details for Banff accommodations well in advance of the meeting. To share ideas or propose collaborations on biologically guided plan optimization, reach out to the conference organisers and join the broader conversation shaping the next wave of clinical practice.