TheraPanacea's CEO says software, not hardware, will scale radiotherapy

Nikos Paragios spent the last days of September 2026 on the exhibition floor of the ASTRO annual meeting in Boston, listening to radiotherapy customers describe the patients they cannot yet treat. Beside him sat Steve Anderson, TheraPanacea's President and GM for the US, who put the question that framed the trip: how a software company answers a capacity problem that machine makers answer with capital equipment. Paragios's bet is that the next capacity gain in radiotherapy comes from software, not a new machine.

Paragios, TheraPanacea's founder and CEO, made the case in a MarketScale Studio segment recorded at the meeting, and he started with the demand side.

The three, in his order, are infrastructure, outcomes and quality of life: enough capacity to treat every patient, better survival once treated, and fewer of the side effects radiotherapy can leave behind. He argues that AI answers each one. Automation raises throughput on the machines a department already has, models that relate past outcomes to treatment choices can recommend options for future patients, and adapting a plan during treatment cuts side effects.

The timing was the meeting itself. ASTRO said in a May 2026 release that its 68th annual meeting would run September 26 to 30 at the Thomas M. Menino Convention and Exhibition Center, draw more than 10,000 attendees, and put artificial intelligence on its panel program beside a storytelling session on building a radiation oncology center in a low-to-middle income country. TheraPanacea's own site dates its AI for planning and adaptive to 2025, six years after ART-Plan's first clearance, and that newer layer is what Paragios was selling in Boston.

The shortfall is counted in machines before it is counted in software

A February 2025 study in The Lancet Oncology put a number on the first of Paragios's three challenges. Fabio Moraes, Brooke Wilson, Gustavo Arruda Viani and colleagues built a Linac Shortage Index, the ratio of demand for linear accelerators to the number installed, where a linear accelerator, or linac, is the machine that generates the treatment beam and 450 patients a year per machine marks the line between enough and too few. With data cut off on September 15, 2024, the global median index was 130, which the authors read as a 30% shortfall in radiotherapy capacity.

Thirty-six of 181 countries have no linear accelerator at all, according to The Lancet Oncology. The spread by income was wide: low-income countries had a median index of 1,523, lower-middle-income countries 399, and high-income countries 96, which sits under the no-shortage line. Of the countries with no machine, 12 were low income and 16 lower-middle income.

That bounds the thesis. Software cannot plan a treatment on a linac that does not exist, and for the countries with none the problem starts there. Paragios's claim is narrower than software replacing hardware: he says throughput gains from automation matter most where infrastructure cannot scale, which in the Lancet grading describes the 105 countries in grades 1 to 3 (arithmetic on the study's counts of 32, 35 and 38) that own linacs but not enough of them.

The Lancet authors drew their cancer counts from the Global Cancer Observatory, the International Agency for Research on Cancer platform that publishes GLOBOCAN. Its 2024 estimates, published by the American Cancer Society and IARC in CA: A Cancer Journal for Clinicians in 2026, cover 34 cancers in 186 countries. Paragios's point for a department head is that every machine on the floor will be asked to treat more people a day before a new one arrives.

What ART-Plan does between the scan and the beam

Every course of radiotherapy starts with a planning scan on which someone outlines the tumor and the healthy structures around it that must be spared, a step called contouring or segmentation. Those spared structures are the organs at risk. The outlines go into a treatment planning system, or TPS, the software that computes beam angles and dose so the tumor receives its prescription while the organs stay under their limits.

ART-Plan, on TheraPanacea's own site, is positioned as an AI assistant to the TPS. The company describes models that run in the background without user supervision, return more than 270 organ and lymph node structures from CT, MR or synthetic CT, and work with any TPS. Each structure follows a published contouring guideline: the site lists 46 head and neck organs drawn to DAHANCA 2020, EPTN 2021 and 2020 international consensus definitions, 19 head and neck lymph node levels drawn to the 2014 Grégoire guideline, and male pelvis organs that follow RTOG guidance from 2012.

On its site TheraPanacea claims ART-Plan cuts manual contouring tasks by more than 93%. The same material covers fusion, the alignment of scans from different modalities, CT, MR, PET and cone beam CT, using rigid registration, which shifts and rotates one image onto another, and deformable registration, which warps the image to follow tissue that moved or shrank. Alignment of that kind is the step that precedes adapting a plan: the new anatomy is mapped onto the old one, then the dose is recomputed.

Paragios described the integration target in the interview as any clinical device available on the market, with the software closing the gap from imaging to treatment and adaptation. The design logic, as the company states it, is to take the scans a department already produces and hand structures and plans back to whichever planning system it already runs, which is what vendor-agnostic means in practice. TheraPanacea's site adds that not every model is available in every market.

The governing permission is the device clearance. TheraPanacea's own timeline records that ART-Plan received CE marking, the European conformity mark required to sell a medical device in the EU, and FDA clearance, the US permission to market a device, for organ-at-risk delineation in 2019. The company's public material lists that 2019 clearance and does not list a separate clearance for adaptive planning or dose prediction, so a buyer should ask which functions in the 2025 planning and adaptive layer carry their own regulatory status.

The machine makers bundled adaptive first

Adaptive radiotherapy means re-planning during a course of treatment that can run for weeks, since tumors shrink, bladders fill differently and patients lose weight, and the plan built on day one no longer fits the body on day fifteen. Imaging Technology News described the hardware-bound version in April 2020: Varian's Ethos uses AI on the treatment machine to take a cone beam CT, the lower-dose scan from the linac's own imager, compare it with the plan, and deliver a fresh adaptive plan inside a typical 15-minute slot from setup to delivery. Varian's Ethos received FDA clearance in February 2020 and was first shown at ASTRO 2019, ITN reported.

The same ITN report gave the throughput numbers that make the automation case. Kevin Moore of the University of California San Diego said the Varian AI planning software produced a full plan in 5 to 20 minutes depending on complexity, against days for the traditional process, and that the center had treated well over 1,000 patients with AI-assisted plans, with a human planner reviewing each one. ITN also reported that RaySearch put machine learning into RayStation 8B in 2018 and that Princess Margaret Cancer Centre in Toronto began using it clinically in 2019, training the algorithm on curated sets of similar plans so it finds the closest prior patients and drafts a plan for the new one.

RaySearch was on the Boston floor too. On its own site the company listed booth 724 for September 27 to 29 and demos of deep learning segmentation, one-click synthetic CT, fast automated replanning and a 2026 version of its RayIntelligence analytics that monitors how AI performs in clinical use. A radiation oncology buyer walking that hall could see the automation Paragios describes offered two ways: inside one vendor's machine and planning stack, or beside whichever stack the department already owns.

Paragios's reply reached back across decades. Innovation in radiology, radiotherapy and robotic surgery, he said, came from better hardware, and the limit of that model is now the calendar and the invoice.

What the clinical evidence shows, and what it does not yet show

The published adaptive evidence so far mostly measures workflow on bundled systems. A study in the Journal of Applied Clinical Medical Physics, published in November 2021 by Mikel Byrne and colleagues at Icon Cancer Centre Wahroonga in Sydney, reported early results for Varian's Ethos online adaptive radiotherapy in prostate cancer on three measures: contouring accuracy, plan quality and treatment time. Those are the yardsticks a physics team would apply to any adaptive tool, including a software-only one.

Physician sign-off on every AI contour and plan stays, Nevine Hanna told CancerNetwork in September 2026. Hanna, lead radiation oncologist at the Thompson Proton Therapy Center, said auto-contouring has cut a task that took hours to minutes, that AI planning replaces iterative manual adjustment by physics staff, and that the added consistency narrows the gap between community practices and academic centers, with the radiation oncologist still responsible for validating contours and confirming treatment intent. What counts as a clinically acceptable automated contour or plan is itself an open question, which Hana Baroudi, Kristy Brock and colleagues at MD Anderson Cancer Center took up in a February 2023 paper in Diagnostics.

TheraPanacea's answer to that question, on its site, is physician acceptance as the test: it says every organ and lymph node model is evaluated continuously for clinical acceptability, with training data from multiple academic centers, and it lists partners including Gustave Roussy, Penn Medicine and Charité. A trial tying ART-Plan to fewer side effects or longer survival is not among the company's published results. Paragios frames outcomes as the next destination, with models that learn which treatment choices produced which outcomes and recommend for the next patient on that record.

For the physicist or department head evaluating a software layer, the operational question is who signs the plan and how acceptance is scored. Hanna's account and TheraPanacea's own material agree on the first: the physician does. The scoring is the part a buyer should ask to see in the center's own data before the first patient.

The buyer's arithmetic between a license and a machine

The Lancet authors modelled future investment cost by country from the shortage index, and that is the hardware side of the ledger: they estimated a grade 3 country needs a median of 52 new linacs by 2045 before any software question arises. The software side is growing on its own. According to Grand View Research in a 2026 report, AI-assisted radiotherapy planning software was a $659.6 million category in 2025, projected to reach $1,815.3 million by 2033 at a 13.6% compound annual growth rate, with cloud deployment holding 68% of 2025 revenue and North America 40.4%.

The operator's case splits by where the department sits. A hospital in a grade 0 or grade 1 country, with machines on the floor and a queue in front of them, is the buyer Paragios describes: the gain is more patients per machine per day, which TheraPanacea's site claims through zero-click segmentation and plans in minutes. A ministry in a grade 4 country is answered by the Lancet study first, with a machine, and by software second.

The test Paragios sets is whether a patient can get the best possible treatment without the center buying heavy hardware, and the signal he and Anderson point to is ART-Plan running on any device already installed, closing the gap from imaging to treatment to adaptation. TheraPanacea's site counts 300 clinical sites on five continents today. The Lancet horizon is 2045; Paragios's claim is that the machines already in service can carry more of the load between now and then.