# TheraPanacea aims to spread best-center radiotherapy quality

By MarketScale Newsroom · Published 2026-11-17 · Healthcare on MarketScale
Canonical: https://www.marketscale.com/industries/healthcare/therapanacea-wants-every-hospital-to-plan-radiotherapy-like-the-best-center-1308824

> TheraPanacea CEO Nikos Paragios wants radiotherapy planning at any hospital to match the best cancer centers, with neuro-oncology the five-year target.

## Key points

- TheraPanacea's two stated objectives: carry best-center radiotherapy quality to every patient on the globe, and identify each patient's best treatment at the start from imaging and clinical history.
- A 2026 Nature Communications trial of 500 patients across five centers found AI-assisted contouring lifted mean Dice overlap from 0.857 to 0.902 and cut median contouring time from 55 to 10 minutes, with less variation between centers.
- Siemens Healthineers' Varian business CE-marked AI contouring for Eclipse in May 2026, so Paragios locates TheraPanacea's edge in its team, hospital partnerships and impact focus rather than in the algorithm alone.

In 2014, two of Nikos Paragios's relatives were diagnosed with brain cancer. He was then, as he is now, a professor of mathematics and computer science at CentraleSupélec, part of Université Paris-Saclay, and he had never worked in radiation oncology. The diagnoses sent him into it.

What he found convinced him that mathematics, computer science and patient data could help physicians treat cancer better, and that conviction became TheraPanacea, where he is now CEO. Speaking in a MarketScale Studio team interview, Paragios set the company two objectives: carry the treatment quality of the best cancer centers to every patient on the globe, and work out at the start of treatment which plan suits each individual. For the physician or physics lead who signs off plans, the stakes are the hours a plan takes and whether its quality changes with the hand that drew it.

The timing matters because the field has moved. In October 2026 CentraleSupélec published an interview with TheraPanacea under the title 'Mathematics in the Service of Radiation Therapy', and a month earlier CancerNetwork ran an interview in which a practicing radiation oncologist argued that automated contouring now narrows the variation between community and academic practice. The question for a buyer is no longer whether software can draw a contour but who delivers it, where, and with what evidence.

> Two of my relatives in the family were diagnosed with brain cancer. At that time, I discovered radiation oncology, and I realized that actually computer science, mathematics, and data could help physicians to better treatment for the patients, and that was my main motivation. — Nikos Paragios, CEO, TheraPanacea

## Contouring is where two hospitals' plans start to diverge

Radiotherapy aims high-energy beams at a tumor while sparing the healthy tissue around it. Before any beam is calculated, a physician outlines the target and every nearby organ at risk, slice by slice, on the patient's CT scan; organs at risk are the healthy structures, such as the heart, lungs or spinal cord, whose dose must stay under a limit. That outlining is called contouring, delineation or segmentation, and it is the step software vendors went after first.

Nevine Hanna, lead radiation oncologist at the Thompson Proton Therapy Center, told CancerNetwork in September 2026 that auto-contouring cuts delineation from hours to minutes. She described software segmenting tumors and organs at risk fast enough that clinicians spend their time on where the radiation should go and what it should spare. According to the same interview, AI then helps the planning step by generating achievable dose distributions and optimizing beam arrangements, work that physics staff otherwise do by iterative manual adjustment.

Dose planning is the second half of the workflow. A planner sets dose targets for the tumor and dose limits for each contoured organ, and the planning software searches for beam angles and intensities that satisfy both, so every change to a contour changes the plan. A May 2026 narrative review in Translational Cancer Research by James Chow of Princess Margaret Cancer Centre in Toronto traces this field from automating those steps toward clinical decision support, which is the stage Paragios's second objective describes.

This is also where variation enters. Two physicians drawing the same organ on the same scan will not draw the same line, and the organ's dose limit is tested against whichever line was drawn. Paragios's first objective is, in practice, a claim about that line.

> The first objective is to propagate the best possible from the best centers to every individual at the globe. — Nikos Paragios, CEO, TheraPanacea

## A 500-patient trial measured what AI assistance does to the gap

Paragios does not cite a trial of TheraPanacea's own tools in the interview, so the evidence that AI-assisted contouring narrows variation between centers comes from trials of other models. The clearest is a prospective multicenter trial published in Nature Communications in 2026, registered as NCT05787522, which tested a deep learning model called iCurveE on 500 thoracic and breast cancer patients across five centers. Thirty-seven physicians contoured eleven thoracic organs at risk three ways: by hand, by accepting the AI output as drawn, and by editing the AI output.

AI-assisted contouring raised mean Dice overlap from 0.857 to 0.902 and cut median contouring time from 55 to 10 minutes. The Dice similarity coefficient scores how much a drawn contour overlaps a reference contour, from 0 for no overlap to 1 for identical. The authors also measured the 95 percent Hausdorff distance, the gap between the worst-matching boundary points after discarding the most extreme 5 percent, and found it fell from 8.01 mm with manual contouring to 5.20 mm with AI assistance, across 2,483 contoured organ sets.

The time figure, which the authors put at an 81.63 percent improvement, is the one that gets quoted. The finding that matters for Paragios's thesis is different: the authors wrote that AI assistance reduced performance variability across the five centers and across physicians of differing expertise, and they framed the result as promoting healthcare equity. Paragios speaks of carrying the best centers' quality outward; by arithmetic on the trial's figures, a manual contour sat 0.045 Dice points below an assisted one on average.

Reviews from 2024 show how actively the method is being measured. Researchers at the Technical University of Munich surveyed deep learning autosegmentation for radiotherapy planning in Strahlentherapie und Onkologie in August 2024, and a group of French centers led from the Institut de Cancérologie de Lorraine published a study of AI contouring for organs at risk and lymph node areas in Radiation Oncology in November 2024. An operator evaluating any vendor's model, TheraPanacea's included, has a template in the Nature Communications design: prospective, multicenter, with manual and assisted arms and a variability measure.

## Picking the treatment at the start from images and history

The second objective is harder to measure. Paragios describes identifying, at the beginning of treatment, the best possible treatment for each individual by combining multimodal data and the patient's clinical history. Multimodal data means more than one kind of input at once: the CT and MR images, the pathology, the laboratory values and the record of what the patient has already been through.

TheraPanacea's team has built that kind of fusion before, outside oncology. An October 2020 CentraleSupélec news item on work that TheraPanacea contributed to, alongside Gustave Roussy, the Paris public hospitals group AP-HP, Inserm and Inria, described an AI trained on CT scans from 478 COVID-19 patients drawn from five cohorts. The model first learned to find characteristic lung lesions on more than 20,000 annotated scan slices, then derived 10 interpretable biomarkers from lesion extent, lung involvement and cardiac preservation, and finally combined those biomarkers with age, sex and clinical data to classify severity and short-term prognosis.

That three-stage chain, image to biomarker to clinical classifier, is the mechanism behind Paragios's second objective, applied to a different disease. The same item quoted Paragios, as president of TheraPanacea, saying clinical deployment was planned for 2021 with GE Healthcare under the Bpifrance-funded AI.Dream project. The work appeared in Medical Image Analysis in October 2020.

TheraPanacea's interview does not name the products it sells, the regulatory clearances they hold, its funding, or the hospitals it works with. What it does give is the logic: data from the best centers trains the model, the model carries that practice elsewhere, and each patient's clinical history steers the plan.

## Siemens and MVision already sell AI contouring, so the edge sits elsewhere

The Varian business of Siemens Healthineers announced in a May 15, 2026 press release that its AI contouring for Eclipse had received CE mark. CE mark is the conformity marking that lets a medical device be sold in the European market, and Eclipse is Varian's treatment planning system, the software in which plans are built. The release describes automatic contouring of more than 200 pre-defined structures from CT and MR images, including organs and lymph nodes, offered as a standardized starting point that clinicians review and refine.

MVision AI, which sells AI contouring under the name Contour+, published a case study in June 2025 on St. Luke's Radiation Oncology Network in Dublin, which it describes as Ireland's largest radiotherapy provider, treating more than 5,500 patients a year across three sites. According to MVision's case study, the network saw time savings of up to 70 percent on head and neck, abdominal and thoracic cases, and plans were ready on average 9.7 working hours sooner, which MVision scaled to roughly 200 patients a year who could potentially start treatment earlier. The network ran Varian and Elekta machines with three different planning systems, so the contouring tool had to fit a mixed fleet, a condition most buyers share.

This is the complication for a founder's pitch. With an incumbent's CE-marked tool and a smaller vendor's documented deployment both on the market, an algorithm that draws organs is not by itself a reason to choose TheraPanacea. Paragios does not claim it is. He locates the difference in three places: the people, the partnerships and the focus on impact.

> We are working with the best hospitals, the best institutions, which allow us to have access to the best possible physicians and clinical outcomes which can help us to develop our solutions. — Nikos Paragios, CEO, TheraPanacea

Partnerships, in his telling, are a data strategy. Working with the best hospitals gives the company access to the best physicians and to clinical outcomes, and outcomes are what a model needs if it is to learn which plan worked rather than which contour looked right. Chow's review places the whole field on that path, from automation toward decision support, which means TheraPanacea is one entrant in a crowded research program rather than a lone breakthrough.

## Access and oversight are the hard parts of democratizing

An April 2025 article in the journal Oncology, published by CancerNetwork, by Viviana Cortiana, Kennedy Itodo, Yan Leyfman and colleagues, sets out why democratizing precision oncology is harder than shipping software. According to the authors, AI in precision oncology requires rigorous validation across diverse populations, because models trained on narrow data overfit and perform poorly on patients unlike the training set, which produces algorithmic bias. They name limited access to technology in resource-constrained settings as a barrier in its own right, and call for interdisciplinary and international collaboration to address it.

Hanna adds the governance limit. In her September 2026 CancerNetwork interview she said physician review remains critical, with the radiation oncologist responsible for validating every contour, confirming treatment intent and signing off on the plan's quality. An AI plan in a regional hospital still needs a qualified physician to approve it, so the technology spreads only as fast as the oversight does.

Cost shows up in the one deployment with numbers. MVision's case study says the Dublin network's purchase of its software was supported by the Friends of St. Luke's charitable body. A hospital that funds AI contouring through a charity is a reminder that the tool's price, and not only its accuracy, decides where it lands.

Paragios's answer runs through his first two differentiators. The international team is how a company whose founder teaches at Paris-Saclay reaches patients it will never meet, and the partnerships with leading institutions are how it gets data that represents them. The 2020 COVID work, with patients pooled from five cohorts and institutions, is the pattern, and the access barrier the Oncology authors describe remains the open item.

## Neuro-oncology is the five-year target

Paragios gives TheraPanacea five years to become a landmark reference in healthcare, with neuro-oncology the named field. Neuro-oncology is the treatment of tumors of the brain and nervous system, and the target has a personal root in the diagnoses that brought him into the specialty. It is also contested ground: the Siemens Healthineers release lists diagnosed brain metastases among the structures its CE-marked tool contours.

The path he describes has two legs, which he calls excellence and impact at scale. First the company's tools are deployed at the best cancer centers, which both proves them and feeds the partnerships. Then they are propagated to every individual on the globe.

Running a company is different from running a lab. Paragios, whose prior experience is mostly academic, says the shift is from solving science problems to working with people and helping them exceed, and that the best part of the job is the people and the problems. The payoff he describes is not a metric.

The signal Paragios points to is deployment at the best cancer centers followed by propagation to every patient. The benchmark the Nature Communications authors set is the one a buyer can hold any vendor to, TheraPanacea included: less variability between centers and between physicians of different expertise, measured prospectively across more than one site.

## Sources

- [A prospective multicenter trial of deep learning auto-segmentation for organs at risk in thoracic radiotherapy](https://www.nature.com/articles/s41467-026-70863-9) (Nature Communications)
- [Artificial intelligence in radiation treatment planning: a narrative review from automation to clinical decision support](https://pmc.ncbi.nlm.nih.gov/articles/PMC13265180/) (Translational Cancer Research)
- [Artificial intelligence contouring in radiotherapy for organs-at-risk and lymph node areas](https://pmc.ncbi.nlm.nih.gov/articles/PMC11580215/) (Radiation Oncology)
- [Deep learning for autosegmentation for radiotherapy treatment planning review](https://pmc.ncbi.nlm.nih.gov/articles/PMC11839850/) (Strahlentherapie und Onkologie)
- [Siemens Healthineers announces CE mark for AI Contouring for Eclipse](https://cancercare.siemens-healthineers.com/newsroom/press-release/siemens-healthineers-announces-ce-mark-for-ai-contouring-for-eclipse-helping) (Siemens Healthineers)

Tags: TheraPanacea, radiation oncology, AI auto-contouring, precision oncology, neuro-oncology, treatment planning

---
Source: MarketScale, https://www.marketscale.com/industries/healthcare/therapanacea-wants-every-hospital-to-plan-radiotherapy-like-the-best-center-1308824. Published for AI indexing and citation; cite the canonical URL. Site guide for agents: https://www.marketscale.com/llms.txt
