AI in Radiotherapy Market (2026-2036)

The growth of this market is mainly driven by the labour content of radiotherapy preparation, since contouring organs at risk and target volumes consumes hours of specialist time per patient and is the step automation reaches first, by adaptive treatment delivered on magnetic resonance guided and cone beam computed tomography guided systems, which requires a plan to be recalculated while the patient lies on the couch and is therefore impossible without automated segmentation, and by a regulatory pathway under which image processing software of this kind is a class II device cleared through a 510(k), which keeps the cost and time of market entry far below the burden carried by therapy hardware. Software accounts for a small fraction of the capital and operating cost of a radiotherapy department, so the growth rate reflects the transfer of clinician hours into licensed software rather than an increase in the number of patients treated.

However, factors such as bundled treatment planning reimbursement that does not rise when preparation is automated, the conversion of a labour cost into a licence cost that only a capacity-constrained department can justify, physician editing of automated contours that erodes the time saving, and the limited value of published accuracy metrics in predicting editing behaviour are expected to restrain the growth of this market to some extent. A department receives the same amount for preparing a plan whether contours were drawn manually over three hours or generated automatically and reviewed in twenty minutes, so a department operating below capacity with salaried staff releases time that generates no additional revenue and the licence becomes a net cost.

A plan begins with the delineation of organs at risk and target volumes on a planning scan, and that delineation is performed slice by slice by a dosimetrist, a therapeutic radiographer or a radiation oncologist, taking from under an hour for a simple pelvic case to several hours for a head and neck case in which dozens of structures have to be separated. Delivery by contrast is machine time measured in minutes per fraction. Automation therefore entered at the contouring step because that is where the hours are, and the first commercial products in this market were segmentation tools sold to reduce the interval between simulation and first treatment rather than to improve the plan. The clinical argument for shortening that interval is independent of the efficiency argument, since tumour progression during the planning window is a recognised problem in rapidly proliferating disease.

Adaptive delivery has made automation a technical requirement rather than an efficiency option. A plan recalculated on the anatomy of the day, whether from a magnetic resonance image acquired on an integrated system or from cone beam computed tomography acquired on a conventional linear accelerator, has to be produced while the patient remains immobilised, which leaves minutes rather than hours for segmentation, dose calculation and plan review. No workflow that depends on manual contouring can meet that constraint, so every adaptive system on the market ships with automated segmentation and automated plan generation as part of the delivery chain. This is the reason the fastest-growing application in this market is tied to hardware purchasing decisions rather than to standalone software purchasing, and it is also the reason adaptive software revenue is difficult to separate from system revenue, which this report addresses in its scope rules.

21 CFR 892.2050, the class II classification for automated radiological image processing software, product code QIH, cleared through the 510(k) pathway and defined as providing automated radiological image processing and analysis tools, a definition that covers artificial intelligence software using nonadaptive machine learning and excludes adaptive artificial intelligence algorithms is the classification most of this software reaches the United States market under, and the exclusion of adaptive artificial intelligence algorithms from that definition is the single most consequential regulatory fact in this market. A model that continues to learn from the data of the institution that deploys it falls outside the classification as written, which is why vendors ship frozen models and why the U.S. Food and Drug Administration draft guidance Predetermined Change Control Plans for Medical Devices, issued on 22 August 2024, which recommends that a marketing submission document the modifications a manufacturer intends to make to a device, the methods for developing, validating and implementing them and an assessment of their impact, so that the individual modifications listed in the plan do not each require a new submission, and which applies to devices reviewed through the 510(k), De Novo and premarket approval pathways matters commercially: the plan is the mechanism through which a manufacturer can update a model on a schedule without returning for a new clearance each time. The U.S. Food and Drug Administration draft guidance Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations, published on 6 January 2025 and carrying non-binding recommendations covering marketing submission contents and lifecycle management across the total product life cycle sits alongside it. Both remain draft and carry non-binding recommendations, so the practical position is that model updating is governed by documentation a manufacturer prepares in anticipation of a policy that is not yet final.

21 CFR 892.5050, the class II classification for a radiation therapy treatment planning system, product code MUJ covers the planning system itself, and the distinction between that classification and the image processing classification separates two commercial positions. A vendor holding a planning system clearance sells the environment in which a plan is created and can embed artificial intelligence functions inside it, while a vendor holding only an image processing clearance sells a component that has to interoperate with somebody else environment. The first position carries the customer relationship and the second depends on it, and that dependence is why independent segmentation vendors have concentrated on supporting every major planning system rather than on building their own.

The European Union Artificial Intelligence Act, which entered into force on 1 August 2024 and became generally applicable on 2 August 2026, and under which high-risk artificial intelligence systems embedded in products regulated by the sectoral legislation listed in Annex I apply from 2 August 2028 following the extension of that transition period by the AI Omnibus, while high-risk use cases in the areas listed in Annex III apply from 2 December 2027 sets the European timetable. Software of this kind is placed on the European market as medical device software under the Medical Device Regulation, which is Annex I sectoral legislation for the purposes of the Act, so the obligations that attach to high-risk systems embedded in regulated products reach radiotherapy software later than they reach standalone high-risk use cases. The extension of that transition reduces the near-term compliance cost for vendors already carrying Medical Device Regulation conformity assessment, and it moves the binding European constraint back to notified body capacity and to clinical evidence requirements under the device regulation rather than forward to the artificial intelligence framework.

The World Health Organization Global Health Observatory indicator for radiotherapy unit density per million population, which counts linear accelerators, Cobalt-60 units, brachytherapy equipment, X-ray generators and ion and proton radiotherapy equipment across both public and private sectors, and which draws on the 2020 to 2021 country survey update together with International Atomic Energy Agency Directory of Radiotherapy Centres data for 2017 to 2021 is the basis on which radiotherapy availability is compared across countries, and the disparity it records between high-income and low-income settings is a disparity in machines and in trained staff together. Automation addresses the second more directly than the first, because software that produces a clinically acceptable contour set without a dosimetrist allows an existing machine to treat more patients, and in settings where the shortage of medical physicists and dosimetrists is more acute than the shortage of linear accelerators this is the constraint that binds. That logic supports cloud deployment and remote planning service models in markets that cannot support an on-premise planning team, and it is the basis on which growth outside the established markets is forecast.