Japan operates roughly 800 radiation therapy centers. That number has grown steadily over the past decade — not explosively, but at the methodical pace that defines Japanese healthcare infrastructure investment. What has changed in the last two years is the pressure accumulating inside those centers: fewer dosimetrists to staff them, an aging patient population driving volume upward, and a national guideline update (JASTRO 2024) that, for the first time, explicitly encourages AI-assisted treatment planning. The market is entering what we would call a structural inflection. Not a revolution. A pivot.
We track this closely because it is our operating environment. What follows is our honest read of the landscape — where the money sits, who controls the incumbent technology layer, what the regulatory path looks like, and where we think the next five years go.
Market Size and the TPS/QA Software Spend
Annual treatment planning system and quality assurance software spend across Japan's ~800 RT centers runs approximately ¥12 billion. That figure aggregates software licensing, annual maintenance contracts, and QA module fees — it excludes hardware (linear accelerators, CT-sims) and clinical staffing. As a software addressable market, ¥12B is meaningful but not vast. Penetration is the question, not total addressable market size.
The TPS layer is heavily concentrated. Varian Eclipse holds roughly 55% share — a dominant position that reflects decades of installed base inertia, strong distributor relationships, and deep integration with Aria OIS at larger academic centers. Elekta Monaco sits at approximately 18%, Philips Pinnacle at 12%, and RaySearch RayStation at roughly 10% and growing. That growth trajectory matters: RayStation's scripting-friendly architecture and active development of DLP (Deep Learning Planning) modules has made it increasingly attractive for institutions willing to invest in customization capability.
The remaining share is thin and fragmented. In-house and academic systems, legacy configurations, and early adopter deployments of newer platforms account for the remainder.
AI Tools: Incumbents and Emerging Players
The major TPS vendors have moved aggressively into AI-assisted planning. Varian's RapidPlan (knowledge-based planning) is installed at a meaningful subset of Eclipse sites in Japan and represents the current practical ceiling for AI-assisted dose prediction at scale. Varian's Ethos adaptive radiotherapy platform layers in AI for online adaptation, though uptake in Japan has been slower than in US markets — capital cost and workflow change management are the primary blockers.
RayStation DLP is the most technically sophisticated AI planning module available commercially today. Fully automated IMRT plan generation via deep learning. The challenge in Japan: DLP adoption requires a degree of institutional technical sophistication — scripts, commissioning rigor, clinical validation workflows — that many community hospitals lack bandwidth for. Academic centers have piloted it. Community uptake is lagging.
Siemens syngo.via provides AI-assisted contouring on the imaging side, feeding into the planning chain upstream. It addresses a real bottleneck but is not a TPS replacement.
Among startups, Accuray's Radixact platform competes at the device level, integrating machine learning for treatment optimization. The startup ecosystem proper — companies building AI software that sits alongside or above existing TPS infrastructure — is less developed in Japan than in the US or China. That gap is part of what we are building into.
Regulatory Pathway: PMDA Classification
AI-medical software in Japan is classified under PMDA's Pharmaceutical and Medical Device Act framework. Three tiers matter here.
Class I (general medical devices) covers software with minimal patient risk — administrative tools, non-diagnostic analytics. Class II (specified controlled medical devices) covers software that informs clinical decisions, including AI-assisted planning tools where a physician or qualified operator remains in the loop. Class III is reserved for software that directly drives high-stakes irreversible decisions without human gate — not the territory most AI planning tools occupy.
Practical implication: an AI planning module that produces a plan for physician and dosimetrist review and approval sits comfortably in Class II territory. The approval timeline for Class II SaMD is typically 12–18 months from submission, assuming adequate clinical evidence and quality management system documentation. Longer if PMDA requests additional data. Shorter — sometimes — if the device meets criteria for expedited review under the 2023 reform provisions.
PMDA is not uniquely conservative by global standards. But it is methodical. Plan for 18 months. Hope for 14.
Reimbursement: The Implicit ROI Problem
Japan's radiation therapy reimbursement runs through K-codes (surgical and procedure codes under the social insurance system). K-codes for IMRT, stereotactic radiotherapy, and related procedures were updated in the 2022 and 2024 revision cycles. None of the current K-code structure explicitly rewards AI-assisted planning. There is no add-on code for “AI plan.” No differential reimbursement for AI versus manual workflow.
This means the economic case for AI in Japanese RT centers is built on implicit ROI: throughput improvement, dosimetrist time recovery, plan quality uplift leading to fewer adaptive fractions, potentially reduced downstream toxicity management. These are real. They are also harder to put on a procurement committee spreadsheet than a direct reimbursement line.
Worth noting: the same reimbursement structure existed in the US before AI-specific codes began appearing in select CPT contexts. Japan will likely follow, but the timeline is uncertain. We operate on the assumption that reimbursement reform is a tailwind 3–5 years out, not a foundation to build the near-term business case on.
Adoption Drivers
Three structural forces are pushing AI adoption regardless of reimbursement dynamics.
Dosimetrist shortage. Japan's dosimetry workforce is undersized relative to the installed base of RT centers. Community hospitals — which account for the majority of RT capacity by site count — are running lean. One or two dosimetrists handling a department's full planning load. When that person is sick or leaves, the center is effectively offline for new plan generation. AI that reduces per-plan labor hours from 4–6 hours to 1–2 hours is not a luxury; it is operational continuity.
JASTRO 2024 guidelines. The Japan Society for Therapeutic Radiology and Oncology updated its clinical practice guidelines in 2024 to explicitly recognize AI-assisted planning as an appropriate component of IMRT workflows. This is significant. In Japanese institutional culture, national society guidance carries substantial weight in procurement decisions. Administrators who were previously skeptical now have cover — and in some cases, a mild obligation — to evaluate AI tools.
Demographic pressure. Japan's population is aging faster than any comparably large economy. Cancer incidence in patients over 70 is rising. RT utilization rates will follow. The centers that do not build capacity today will face volume they cannot absorb within this decade.
Adoption Blockers
Honest assessment requires naming the real friction.
Legacy OIS integration is the first blocker. Most Japanese community hospitals run Varian Aria or Elekta Mosaiq as their oncology information systems. Any AI planning tool needs to fit into that workflow — importing and exporting in DICOM RT format, integrating with existing plan approval chains, working within the OIS's user permission model. Integration work that requires custom development per site is a sales and implementation cost multiplier. We have firsthand data on how variable this gets across hospital IT configurations in the Tohoku region.
Physician skepticism — specifically among radiation oncologists who trained on manual planning and carry deep intuitions about what a “good plan” looks like — is the second blocker. This is not irrational. It reflects a reasonable question: does the AI plan meet my clinical standards, or does it optimize for metrics that don't fully capture what I care about? The answer requires strong clinical validation data, which takes time to accumulate.
PMDA timelines are the third. For a startup entering the market, 18-month approval windows mean capital efficiency planning that few early-stage teams do cleanly. It also creates a competitive dynamic where incumbent vendors with existing approvals have a structural advantage that is hard to overcome on speed alone.
2026–2030 Forecast
Our working model: AI-assisted plan share in Japan's RT centers moves from approximately 15% of IMRT plans (2024 baseline) to somewhere in the 45–55% range by 2030. The range reflects genuine uncertainty around the pace of community hospital adoption and the degree to which reimbursement reform accelerates or lags the structural drivers.
The US trajectory — which moved from roughly 20% AI-assisted planning in 2020 to over 60% in 2024 — is not a direct template. Japan differs on all of the dimensions that matter: Eclipse dominance is more entrenched, the venture ecosystem is thinner, the regulatory clock runs longer, and the physician culture around plan ownership is more conservative. But the underlying clinical logic is identical, and the workforce math is if anything more acute in Japan.
The Japan-specific factors that distinguish this market from EU adoption curves are also worth naming. European markets have a more distributed TPS vendor landscape — RayStation and Monaco hold larger shares in Germany and the UK than in Japan — which creates more natural entry points for AI tools that partner with non-dominant vendors. Japan's Eclipse concentration means any AI layer that doesn't integrate cleanly with Aria/Eclipse faces a steeper uphill. It is not insurmountable. But it is not a small thing.
Where This Leaves Us
I am not going to use this space to make a product pitch. The facts of the market landscape are what they are regardless of what Airato does with them. But I will say this: the centers we talk to in Sendai, Morioka, and Fukushima are not waiting for a perfect AI planning tool. They are waiting for one that is good enough, integrates without a multi-month IT project, is approved through PMDA, and has local clinical support they can actually reach.
The market is real. The friction is real. The dosimetrist shortage is not going away. And the inflection — however gradual by US startup standards — is underway.
