© The Authors, published by EDP Sciences, 2026

Introduction

Radiotherapy has been transformed twice in the past quarter-century: first by image-guided and intensity-modulated delivery, and second by the integration of stereotactic body radiotherapy into mainstream practice. A third wave is now arriving, characterised less by a single dominant technology than by the simultaneous maturation of several distinct innovations. FLASH radiotherapy promises a transformation in normal-tissue tolerance through ultra-high dose-rate delivery [1, 2]. Adaptive radiotherapy with magnetic resonance or cone-beam computed tomography guidance enables daily plan re-optimisation [3–5]. Artificial intelligence (AI) is reshaping the planning workflow [6, 7]. Theranostics with targeted radionuclide therapy is delivering Phase III-validated systemic radiation [8–10]. Spatially fractionated radiotherapy (SFRT) using GRID and LATTICE techniques has re-emerged as a strategy for bulky and radioresistant tumours [11–13]. And particle therapy continues to expand its biologically rationalised footprint [14]. Video discussing these technological innovations available at Supplementary Video 1.

Each of these innovations has been the subject of extensive recent review. What is less often addressed, and what this article seeks to provide, is a unified framework for comparing them – not on biological promise alone, but on the joint axes of evidence and capital intensity. A second point that is easily lost in modality-by-modality reviews is that these innovations are not wholly independent of one another. Several stand in enabling or prerequisite relationships – FLASH presently depends on particle infrastructure for most non-superficial targets, and online adaptive radiotherapy depends on AI-based contouring and plan evaluation to be clinically practical – so that the diffusion of one innovation is partly governed by the diffusion of another. The framework, therefore, treats interdependence as an explicit axis of analysis rather than an afterthought. The central argument is that capital cost, more than any other variable, will determine which innovations actually reach patients around the world in the next five years. This review proceeds in four parts: the comparative framework itself; an impact ranking integrating evidence, reach, and capital; a deeper analysis of capital intensity as the dominant diffusion variable; and the implications for training, trial design, and health-system policy.

A four-axis comparative framework

The six innovations are mapped in Table 1 across four axes: physical principle, biological rationale, current evidence level, and dominant adoption barrier. The evidence axis uses a five-tier scale: Tier 1, preclinical or first-in-human; Tier 2, Phase I–II single-institution; Tier 3, multi-institutional Phase II; Tier 4, Phase III or randomised controlled trial data; Tier 5, guideline-incorporated standard of care.

Evidence gradient

Three observations emerge from the matrix. The first is that the evidence gradient is uneven. Theranostics and particle therapy sit at Tier 5 for specific indications – lutetium-177 prostate-specific membrane antigen (PSMA) for metastatic castration-resistant prostate cancer [9] and lutetium-177 DOTATATE for neuroendocrine tumours [10] are both guideline-incorporated, while protons remain standard for paediatric, skull-base, ocular, and chordoma indications, a narrow spectrum of clinical indications. Adaptive radiotherapy has reached Tier 4 with the SMART pancreas trial [3] and MIRAGE prostate trial [4]. AI sits at Tier 3 for workflow endpoints but lacks Tier 4 outcome data. FLASH and SFRT remain at Tier 2 despite substantial preclinical and early-clinical interest. The gap between Tier 2 and Tier 5 spans most of the modern history of evidence-based radiation oncology, and yet – as the impact ranking will show – evidence quality is not the principal determinant of diffusion.

Convergence on radiation immunology

The second observation is that the biological rationales converge on a shared theme. Five of the six innovations explicitly invoke immune mechanisms: the FLASH effect preserves circulating lymphocytes and reduces normal-tissue oxygen depletion [1]; adaptive radiotherapy reduces integral dose to lymphoid organs; SFRT generates bystander and abscopal responses through valley-region preservation of immune cells [12]; theranostics produces immunogenic systemic radiation; and particle therapy generates emerging immunogenic data. AI is the only exception, and only because its biological link is indirect, mediated through radiomics-based prediction of biological response [6]. This convergence suggests that radiation immunology may be the unifying scientific narrative of the coming decade, and that combination strategies pairing these modalities with checkpoint inhibitors or cellular therapies will be a major area of trial design. The implication for cancer biologists is that the boundary between radiotherapy research and tumour immunology research is dissolving, and combined-modality investigators – not single-modality specialists – will increasingly define the field.

Two families of physical principles

The third observation is that the physical principles cluster into two distinct families. FLASH, adaptive radiotherapy, SFRT, and particle therapy are delivery innovations that change how dose is shaped in space or time at the linear accelerator. AI and theranostics are biological or computational innovations: AI changes how plans are derived from imaging, and theranostics changes the route by which radiation reaches tumour cells. This distinction is operationally important because the two families have different capital structures – delivery innovations typically require new hardware, while biological and computational innovations layer onto existing infrastructure. The implication for departmental planning is that the latter family will scale more rapidly than the former, almost regardless of the underlying biology.

Interdependence among the innovations

A fourth observation, less often made explicit, is that these innovations are coupled. Three couplings are worth naming. First, FLASH radiotherapy for most non-superficial tumours currently requires particle delivery: with the exception of superficial targets reachable by electron FLASH, the only platforms able to deliver ultra-high dose rates at depth in the near term are proton – and potentially carbon-ion – systems. Human trials of FLASH for deep-seated disease over the next five years will therefore be conducted almost exclusively within particle centres, while the field awaits very-high-energy electron (VHEE) and ultra-high-dose-rate X-ray systems. Particle infrastructure is, in this sense, a practical prerequisite for clinical FLASH, and the diffusion of the two innovations is partly linked. Second, online adaptive radiotherapy depends on AI: daily plan re-optimisation is only clinically feasible because deep-learning auto-contouring and automated plan evaluation compress what would otherwise be a multi-hour manual task into minutes. Advances in AI-based segmentation and plan checking are thus a rate-determining input to the adoption of real-time adaptive workflows, whether MR- or CT-guided. Third, SFRT and FLASH converge within particle centres, where proton pencil-beam scanning can deliver lattice dose distributions at FLASH dose rates. These couplings mean that the population-level impact of any single innovation cannot be read off its own evidence and capital profile alone; where an innovation enables or gates another, its effective reach is larger or smaller than it first appears. The rankings that follow are presented for each innovation individually, but the interdependences above should be read alongside them.

Ranking by anticipated impact

Table 2 ranks the innovations by anticipated population-level impact over five years, integrating three sub-criteria: speed of adoption, breadth of patient reach, and capital intensity. The resulting order departs sharply from a ranking based on evidence alone. It is essential to read this ranking as a forecast of diffusion and accessibility, not as a ranking of intrinsic biological or clinical efficacy. A modality may rank low because its capital pathway is steep, while remaining, for the selected patients it serves, the most effective treatment available – carbon-ion therapy for radioresistant histologies is the clearest example. The two questions – how much benefit does this innovation deliver to an eligible patient, and to how many patients will it realistically become available within five years – are deliberately kept separate throughout what follows.

AI in planning: the universal layer

AI in planning and automation occupies the first rank. The reasoning is straightforward. Auto-segmentation tools, dose-prediction models, and synthetic computed tomography (CT) generation are deliverable as software upgrades for any existing linear accelerator [7]. With approximately 20,000 linear accelerators worldwide, the scaling potential is unmatched. The bottleneck is regulatory clarity for adaptive learning models and the demonstration of clinical – rather than purely workflow – endpoints, not infrastructure. Several auto-contouring products already hold regulatory clearance in major jurisdictions; the next five years will see the contest move from feasibility to outcome studies, with particular attention to whether AI-assisted planning reduces inter-planner variability sufficiently to translate into measurable differences in toxicity or local control.

Theranostics: a guideline-incorporated systemic radiation

Theranostics is ranked second. Lutetium-177 PSMA and lutetium-177 DOTATATE are already guideline-incorporated for prostate and neuroendocrine indications affecting hundreds of thousands of patients annually [9, 10]. Diffusion is constrained by isotope supply, particularly for actinium-225, and by nuclear medicine infrastructure rather than by hospital capital outlay. The major institutional cost is consumable, not capital. Several emerging targets – fibroblast activation protein, human epidermal growth factor receptor 2, carbonic anhydrase IX – are entering clinical trials, suggesting that the theranostics platform will broaden well beyond the prostate and neuroendocrine indications that opened the field. The training implication is that radiation oncology and nuclear medicine, traditionally distinct training pathways, must develop integrated dosimetry expertise.

SFRT: zero-capital diffusion

SFRT is ranked third – a placement that may surprise readers given the absence of Tier 3 or 4 evidence. The justification is its uniquely low capital barrier. SFRT requires no new equipment beyond existing multileaf collimators or modest 3D-printed collimator inserts that cost less than US$5,000 [11]. With strong rationale for bulky and radioresistant tumours and an emerging multi-institutional trial portfolio [13], SFRT has a clear path to global diffusion that is paced only by training and evidence generation, not by financing. The convergence of SFRT with FLASH – proton pencil-beam scanning ultra-high-dose-rate three-dimensional lattice radiotherapy has now been demonstrated as proof-of-concept, suggesting that within proton centres, the two innovations may eventually be deployed jointly rather than sequentially.

Adaptive radiotherapy: capital-constrained despite Phase III evidence

Adaptive radiotherapy is ranked fourth despite its Tier 4 evidence. The central limitation is capital. Magnetic resonance linear accelerators (MR-Linacs) cost more than US$7.5 million for acquisition alone [15], require magnetic-shielded vaults, and treat fewer patients per day than conventional systems – typically five to six versus thirty to forty. CT-guided online adaptive systems at lower cost represent the more realistic mass-diffusion path. The MIRAGE trial demonstrated meaningful toxicity reduction in prostate stereotactic body radiotherapy with MR guidance compared with CT guidance [4], but the absolute toxicity differences must be weighed against the capital opportunity cost: the same investment could deploy AI-assisted planning across hundreds of conventional linacs. The realistic diffusion of online adaptive radiotherapy is, moreover, tied to progress in AI: daily re-planning becomes clinically tractable only when auto-contouring and automated plan evaluation are fast and reliable enough to be completed with the patient on the couch. Adaptive radiotherapy and AI-assisted planning should therefore be understood as complementary rather than competing investments, with the former partly enabled by the latter.

FLASH: biology ahead of evidence

FLASH radiotherapy is ranked fifth. The biology is compelling, and the FAST-01 and FAST-02 trials have established feasibility [1, 2], but Phase III data are absent, and the dose-rate hardware modifications are non-trivial. Initial diffusion will be in centres already operating proton therapy systems capable of FLASH dose rates. As noted in the interdependence analysis above, this concentration is not incidental: for non-superficial targets, particle infrastructure is presently a prerequisite for clinical FLASH, so FLASH diffusion over the planning horizon is gated by the same capital barrier that constrains ranking six. Electron FLASH may reach broader use sooner, but is restricted to superficial targets. The fundamental radiobiology of the FLASH effect – in particular the role of transient oxygen depletion and the contribution of immune-cell preservation – remains incompletely characterised, and the field would benefit from coordinated multi-institutional preclinical work alongside the early clinical trials.

Particle therapy: high evidence, low absolute reach

Particle therapy is ranked sixth in absolute population reach despite its Tier 5 evidence for niche indications. The capital intensity is the defining barrier. Proton-only facilities cost approximately US$95 million; combined carbon-ion and proton facilities exceed US$138 million [14]. These headline figures, however, describe large multi-room facilities (typically four or more treatment rooms, inclusive of construction, shielding, and in many cases land acquisition) and should not be read as the cost of entry. Compact single-room proton systems now in routine use are reported at roughly US$30–50 million [16, 17], a materially different capital proposition, and the figure is not directly comparable to the single-machine, equipment-only MR-Linac acquisition cost cited above. The lower capital barrier of single-room systems does not, however, necessarily translate into a lower cost per unit of patient throughput. A single treatment room serves far fewer patients per year than a multi-room facility sharing one accelerator across several rooms, so fixed costs – accelerator, building, staffing, and quality assurance – are spread over a smaller caseload, and the cost per fraction or per patient can be higher than at a well-utilised multi-room centre. The single-room advantage is therefore one of accessibility and entry cost rather than of unit efficiency. Newer designs that eliminate the rotating gantry – fixed-beam or compact mount-less configurations – can nonetheless reduce both equipment and building cost materially, since the gantry and its shielding and footprint are among the largest cost drivers, and represent one of the more promising routes to lowering the entry barrier further. Fewer than 130 proton centres and approximately 15 carbon-ion centres exist globally. Compact single-room proton systems in the US$30–40 million range may modestly erode this barrier, but will not democratise access across the globe, but rather reflect existing inequalities in access to capital and governmental funds. A purely per-modality capital figure can also mislead in a second way: within a department, a single-room proton system frequently functions as an anchoring technology that raises the case-mix, referral base, and reputation of the whole centre, so that its value is not fully captured by evaluating it on its standalone cost. Where such ecosystem effects are material, the capital case should be made at the level of the centre rather than the machine. Within the niche indications where particle therapy is biologically superior – paediatric tumours, chordoma, ocular melanoma, and selected re-irradiation cases – the modality will remain standard of care; outside those indications, randomised data are mixed, and the case for displacement of photon therapy is weaker than the marketing literature suggests.

The spread between highest and lowest ranks

The spread between the highest- and lowest-ranked innovations is instructive. AI-assisted planning could be operating on every linear accelerator within five years; particle therapy is unlikely to expand beyond a small fraction of countries within the same horizon. This is not a value judgement on the underlying biology – carbon-ion therapy may well be the single most biologically advantageous modality for radioresistant histologies such as recurrent rectal cancer or chordoma. It is a statement about diffusion. The question for departmental leaders is therefore not which innovation is most exciting, but which combination of innovations is realistically deployable within the planning horizon and the financing envelope. For most centres, the realistic combination over the next five years is: AI-assisted planning as the universal layer, theranostics referral pathways for metastatic disease, SFRT for bulky and re-irradiation cases, and – where capital permits – selective investment in CT-guided adaptive radiotherapy. FLASH and particle therapy will remain referral-only for the great majority of centres.

Capital cost is the dominant diffusion variable

Table 3 disaggregates the capital structure of each innovation into equipment cost, building or facility cost, and a qualitative capital band. Two cautions apply to every figure in that table. First, the costs are approximate, drawn from published cost-analysis studies, and vary substantially by region, vendor, configuration, and year; they are intended to convey order-of-magnitude differences between capital bands, not precise budgeting estimates. Second, the figures are not all like-for-like: the particle-therapy headline cost reflects a multi-room facility inclusive of construction and land, whereas the MR-Linac and CT-adaptive figures reflect single-machine equipment acquisition. The comparison that matters is between capital bands and between the structural requirements (new accelerator, shielded vault, dedicated footprint versus none), not between individual dollar figures read in isolation. The pattern is striking. The three highest-impact innovations – AI, theranostics, and SFRT – share a common feature: they layer onto existing infrastructure. AI runs on the same linacs already installed; SFRT delivers on those same linacs with no new hardware; theranostics deploys through existing nuclear medicine departments with modest hot-lab augmentation. The three lowest-impact innovations – adaptive radiotherapy with MR guidance, FLASH, and particle therapy – each require either new accelerators, magnetic-shielded vaults, or dedicated facility footprints measured in thousands of square meters.

Equity implications

This pattern carries direct equity implications. Radiotherapy access is already deeply uneven: more than half the world’s population lacks access to basic radiotherapy services [18]. Innovations that depend on new high-capital infrastructure cannot close this gap; they may widen it. Innovations that layer onto existing linear accelerators offer the only realistic mechanism for global access to advanced radiotherapy within a five-year horizon. AI-assisted planning could democratise plan quality. SFRT could extend treatment to bulky tumours in centres without proton capability. Theranostics could deliver systemic radiation to metastatic disease through nuclear medicine infrastructure that is more widely distributed than radiation oncology bunkers. The contrast with the equity-neutral framing common in technology-launch communications is stark, and the field would benefit from explicit equity-impact statements accompanying major capital investments.

Capital as an ecosystem rather than a line item

A purely per-modality view of capital, while analytically convenient, can distort departmental decision-making in a way that deserves explicit acknowledgement. High-capital platforms rarely sit in isolation; they reshape the clinical and economic ecosystem around them. A single-room proton system, for example, may function less as a standalone cost centre than as an anchoring technology – drawing complex referrals, supporting clinical trials, justifying subspecialty recruitment, and raising the case-mix and reputation of the entire centre – so that its contribution to institutional value is understated by any analysis that evaluates it on its own throughput and debt service alone. The same logic applies, more weakly, to MR-Linacs as flagship adaptive platforms. None of this overturns the central thesis that capital intensity governs diffusion; the great majority of centres will still be unable to acquire these platforms at all. But for the minority of centres weighing such an investment, the appropriate unit of analysis is the centre, not the machine, and the cost case should be built on the incremental value to the whole programme rather than on the modality’s standalone economics. Recognising this does not dissolve the equity problem – it sharpens it, because the ecosystem benefits of anchoring technologies accrue precisely to the well-resourced centres already best placed to afford them.

Patient selection distortion

Persistent capital cost has a second consequence, less often discussed: the distortion of patient selection. Centres that have invested US$95 million in a proton facility face strong financial pressure to maintain throughput, which can bias selection toward indications that generate revenue rather than indications where proton therapy offers the greatest biological advantage. Similar pressures apply to MR-Linacs. Independent referral pathways, cooperative-group trial enrolment requirements, and transparent indication criteria are partial mitigations, but the underlying tension between capital servicing and evidence-based selection is structural [19]. The radiation oncology community has, on balance, addressed this tension less directly than the surgical oncology community, where high-volume centre arguments and credentialing standards are routine. Adopting analogous mechanisms – minimum case volumes for proton centres, mandatory enrolment in registry trials, transparent reporting of indication mix – would help to distinguish appropriate capital concentration from inappropriate self-referral.

Operational cost trajectories

A third dimension worth considering is the operational cost trajectory after capital is amortised. For particle therapy facilities, approximately 87% of annual costs are operational rather than capital-amortisation [14]; for photon facilities, the figure is 92%. This means that even when capital is depreciated over a 30-year facility lifespan, the staffing, maintenance, beam-line consumables, and quality-assurance burden of particle facilities remains substantially higher than for photon centres. Operational cost is a weaker but persistent constraint on diffusion that will not be solved by capital subsidy alone. Health-system planners contemplating new particle-therapy investments should model 30-year operational envelopes, not merely up-front construction costs, and should evaluate whether projected patient volumes can sustain those envelopes without distorting referral.

Policy levers

Policy levers that could partially reshape this landscape fall into four categories. First, national capital subsidy programmes for hub-and-spoke proton networks, of the kind adopted in the United Kingdom and Australia, can extend access without each centre bearing full capital cost. Second, vendor leasing or pay-per-use models can spread capital over operating budgets, particularly for MR-Linacs and compact proton systems. Third, regional cooperative consortia can share single high-cost facilities across multiple referring institutions, with centralised credentialing and quality assurance. Fourth, tiered-pricing software contracts – already common in radiology AI – can enable lower-resource centres to access advanced planning tools at a fraction of the cost. Each of these mechanisms has been demonstrated in practice; none has yet been deployed at sufficient scale to close the global access gap. The collective failure of the field to scale these mechanisms is one of the most consequential implementation gaps in twenty-first-century cancer care.

Implications for trial design and training

The framework presented here has implications beyond institutional capital allocation. For trial design, the asymmetry between high-evidence and low-diffusion innovations argues for differentiated trial endpoints. Innovations with high capital intensity and limited reach – particle therapy, MR-guided adaptive radiotherapy – should be evaluated against pragmatic comparators using cost-effectiveness as a co-primary endpoint, not merely toxicity or local control. Innovations with low capital intensity – AI, SFRT – should be evaluated through cluster-randomised or stepped-wedge designs that test whether broad deployment achieves population-level benefit, rather than through traditional individual-patient randomisation in tertiary centres. Cooperative groups have begun to adopt these designs in radiotherapy contexts, but at insufficient scale, and the next five years will require explicit prioritisation of pragmatic and implementation-science endpoints alongside conventional efficacy endpoints.

For training, the implication is more direct. Postgraduate radiation oncology curricula are still organised around delivery technologies that residents will use throughout their careers, but the framework presented here suggests that the technology that will most reshape day-to-day practice for the average graduate is AI-assisted planning, followed by theranostics referrals and SFRT. Curricula that allocate disproportionate teaching time to particle therapy and MR-guided adaptive radiotherapy may be over-investing in technologies that the majority of trainees will never operate, while under-investing in technologies they will use weekly. Realigning training emphasis with realistic deployment trajectories is a tractable institutional response to the diffusion asymmetry described above. Specific curricular changes worth considering include dedicated AI and radiomics modules, joint radiation oncology/nuclear medicine theranostics rotations, and SFRT planning workshops integrated into physics training.

Limitations of this framework

Three caveats apply. First, the framework treats each innovation as a class. Within any class, individual indications occupy different evidence tiers – lutetium-177 PSMA is guideline-incorporated, whereas actinium-225 PSMA remains in early trials; protons are standard for paediatric central nervous system tumours but unproven for adult prostate cancer outside trial settings. Second, capital cost figures are approximate and drawn from published cost-analysis studies; actual expenditure varies by region, vendor, and configuration, and – as emphasised in the capital section – headline particle-therapy figures describe multi-room facilities and are not directly comparable to single-machine equipment costs. Third, the five-year adoption-speed estimates assume current trajectories of evidence generation, regulatory approval, and reimbursement, and could shift with policy or supply-chain changes – most notably for actinium-225, where global production capacity is the rate-limiting variable. A fourth caveat follows from the analysis above: because several innovations are interdependent – FLASH on particle infrastructure, online adaptive radiotherapy on AI – the rankings are not fully separable, and a shift in the diffusion of one innovation will move others. The single-innovation rankings should therefore be read together with the interdependence analysis rather than as independent forecasts. Finally, the framework deliberately ranks expected diffusion rather than intrinsic efficacy; a low rank should never be read as a recommendation against a modality for the individual patients in whom it is clinically superior. The framework should therefore be read as a structured starting point for institutional discussion rather than a definitive ranking, and updated as the underlying evidence and infrastructure evolve.

Conclusion

Six innovations will reshape radiation oncology over the next five years, but they will not reach patients equally. AI-assisted planning, theranostics, and spatially fractionated radiotherapy are positioned to deliver the largest absolute population benefit because they bypass the capital barrier that constrains MR-Linacs, FLASH systems, and particle therapy. Capital cost is therefore not merely an institutional planning consideration but the principal determinant of global health equity in radiation oncology. Trainees, departmental leaders, and health-system planners would be served by reading future technology announcements through this lens – asking not only whether the biology is compelling, but what the capital pathway to bedside is, and who pays for it. Subsidy frameworks, regional consortium models, vendor leasing arrangements, and tiered-pricing software contracts may partially mitigate this asymmetry, but they are unlikely to close the gap within the planning horizon. Recognising the asymmetry between biological excitement and infrastructural reality is the first step toward a more equitable next decade in radiation oncology, and toward a research and training agenda that takes diffusion as seriously as it takes discovery. The framework presented here is offered as a structured starting point for that work – a way of asking the right questions about each new technology rather than a definitive answer about any of them. The questions are simple enough to be applied at any departmental meeting: What is the physical principle? What is the biological rationale? What is the evidence level today, and Where is it heading? What is the capital intensity, and what is the dominant adoption barrier? Which other innovations does it enable or depend upon? Asked consistently, these questions discipline our enthusiasm, surface our blind spots, and help to align institutional investment with realistic patient benefit.

Funding

This research received no external funding.

Conflicts of interest

The author certifies that he or she has no financial conflict of interest (e.g., consultancies, stock ownership, equity interest, patent or licensing arrangements) in connection with this article.

Data availability statement

This article has no associated data generated and/or analysed.

Ethics approval

Ethical approval was not required.

Informed consent

This article does not contain any studies involving human subjects.

Supplementary material

Supplementary Video 1: Radiation oncology, five years and beyond. Access Supplementary Material

Acknowledgments

The author acknowledges discussions with colleagues that informed the framework presented here. An AI assistant (Claude, Anthropic) was used to assist with literature triage, and table layout and writing; all scientific content, interpretation, evidence synthesis, and conclusions are the responsibility of the author, who has reviewed and verified each citation.

References

All Tables