Radiotherapy auto-contouring & AI treatment planning

Preparing a radiotherapy plan means outlining dozens of organs and the tumour slice by slice, then iterating dose optimisation until the physicist is satisfied. AI now performs the contouring in minutes, predicts the achievable dose, produces a first plan automatically and, on adaptive machines, redoes it for every fraction. Clinics report large reductions in planning time and less inter-observer variation; oncologist and physicist still review and approve every contour and plan.

Problem & value

• Today: radiation oncologists, dosimetrists and physicists manually delineate organs at risk (OARs), target volumes and nodal regions on planning CT/MR, then plan iteratively in the treatment planning system (TPS). It is a bottleneck: contouring is slow, varies between observers and is the main barrier to online adaptive radiotherapy, where the patient lies on the table while the plan is redone.
• AI value: (1) consistent, guideline-conformant OAR and target proposals that only need editing; (2) dose prediction that tells the planner what is achievable before optimisation; (3) automatic plan generation that reduces plan-quality variance between centres; (4) fast re-contouring and replanning for adaptive workflows; (5) synthetic CT enabling MR-only planning.
• Evidence: the vendor and clinical literature (e.g. ESTRO-linked publications, MVision, Therapanacea, Duke auto-planning work) report planning times falling from roughly 45-120 minutes to 6-15 minutes for routine sites; figures vary by site and are vendor/centre-reported, so measure locally [1][3][4].

Variants

• Auto-contouring only: OARs for head and neck, thorax, pelvis, breast; the entry point for most clinics.
• Target-volume and nodal auto-contouring: CTV/nodal regions per ESTRO or RTOG consensus; needs more editing.
• Dose prediction and knowledge-based planning: 3D dose or DVH prediction to guide or warm-start optimisation.
• Fully automated planning: scripted or ML-driven IMRT/VMAT optimisation (e.g. Eclipse RapidPlan, RayStation machine-learning planning, Duke-style auto-planning) with planner review.
• Online adaptive RT: Varian Ethos (CBCT) and MR-Linacs (Elekta Unity, ViewRay MRIdian) with daily AI contouring and replanning.
• Synthetic CT: MR-to-sCT and CBCT-to-sCT for MR-only workflows and dose calculation on daily images.
• Particle therapy: proton planning adds range-uncertainty and robust optimisation constraints.
• Veterinary / research: non-clinical contouring with open tools, outside MDR scope.

How it works

1. planning CT/MR (DICOM)
2. pre-processing (orientation, protocol/site detection)
3. 3D segmentation (OARs, GTV/CTV, nodes)
4. DICOM-RT Struct into TPS
5. [HUMAN] oncologist reviews/edits contours, defines targets & prescription
6. dose prediction (3D U-Net / GAN-style)
7. plan optimisation (IMRT/VMAT)
8. [HUMAN] physicist/dosimetrist plan review, independent dose check, patient-specific QA
9. treatment; adaptive: daily CBCT/MR
10. sCT, re-contour, re-optimise
11. [HUMAN] approve on-couch

• Segmentation: 3D convolutional or transformer networks (nnU-Net-style) trained on expert contours following consensus guidelines; output as RT Structure Sets.
• Dose prediction: image-to-image networks conditioned on contours and beam geometry predict the dose grid; the deliverable plan is still computed by the TPS dose engine, not by the network.
• Plan optimisation: classical inverse-planning optimisers, guided by ML priors or reinforcement-learning and rule-based "wish list" automation.
• Human in the loop: every contour and plan is approved by a qualified oncologist and physicist; automation bias and edge anatomy (post-surgical, implants, paediatric) are the main review risks.

Data preparation

Radiotherapy AI runs on DICOM data that already lives in the PACS, the oncology information system and the treatment planning system (TPS); paper and file shares play almost no role. Typical gaps are inconsistent structure names (the same organ labelled five ways), contours that follow different guidelines over the years, missing prescriptions in structured form, and a mix of DICOM-RT versions from different vendors. Access rights are usually fine-grained but tied to clinical roles.

Preparation steps: (1) inventory the structure sets and agree on one naming standard (TG-263) and one contouring guideline (ESTRO, RTOG) per site; (2) map the old names to the standard with a script, for example with pydicom or the TPS scripting API; (3) select 20 to 50 recent, clinically approved cases per site as a local validation set, reviewed by physician and physicist; (4) pseudonymise any data leaving the clinic network, see PREP-022 and PREP-018 for building test sets with domain experts; (5) check the interfaces between scanner, PACS and TPS.

Rough share of effort: about 30 to 40 percent of the project, mostly clinician time for naming clean-up and validation cases. Use PREP-002, PREP-018 and PREP-019 as models. See also GUIDE-003.

Implementation paths

Which path for whom

• Enterprise (hospital groups, MVZ, radiotherapy networks): Path A, because auto-contouring and planning modules are MDR-certified products that plug into the treatment planning system and the vendor carries the conformity burden.
• Developer: Path C, because open models (TotalSegmentator, nnU-Net, MONAI) are suited for research, benchmarking and validation, not for clinical release without own MDR responsibility.
• Public (university hospitals, municipal clinics): Path A with an EU-hosted or on-premises deployment, because procurement rules and patient data favour certified products inside the clinic network.

Path A - Buy (recommended)

• Auto-contouring: MVision AI (Finland, EU), Therapanacea ART-Plan (France, EU), Limbus AI (UK), RaySearch RayStation deep-learning segmentation (Sweden, EU), Varian/Siemens Healthineers (Eclipse, Ethos, AI-Rad Companion Organs RT), Elekta ADMIRE/Mosaiq, Brainlab Elements, Philips, DirectORGANS-class tools.
• Planning/adaptive: Varian Ethos and RapidPlan, RayStation, Elekta Unity, ViewRay MRIdian, Accuray.
• Choose A because the product is the MDR-certified medical device (typically class IIa/IIb) with vendor-held clinical evaluation and TPS integration through DICOM-RT.

Path B - Build on cloud APIs

• Little role for general LLM APIs in the core. Cloud-hosted vendor segmentation services (MVision, Therapanacea) are consumed via DICOM transfer; general VLMs are unsuitable for contour generation.
• LLMs can help with surrounding tasks: protocol look-up, plan-report drafting, QA-log summaries (HLTH-004), but not with the dose-relevant outputs.

Path C - Self-host (open source)

• Models/tools: nnU-Net, MONAI, TotalSegmentator (CT/MR), MedSAM/SAM-Med3D-style foundation models, OpenKBP-derived dose prediction baselines, matRad and OpenTPS for research planning, pydicom/rt-utils for DICOM-RT.
• Hardware: HW3 (one 48 GB GPU) for training and inference; inference alone runs on HW2.
• Caveat: in-house clinical use makes the hospital the manufacturer under MDR Art. 5(5) health-institution rules (in-house exemption: no equivalent device on the market, QMS, justification, public declaration). Research use needs ethics approval and no clinical decision depends on it.

Path D - Custom training

Worth it for institution-specific contouring styles, rare sites, paediatric cases or MR-Linac protocols. Needs a few hundred to thousands of curated structure sets with consistent nomenclature (TG-263), multi-observer review and a prospective validation. Federated training across centres helps with privacy (DATA-005).

State of the art (2026)

• Auto-contouring of OARs is routine in many European clinics; accuracy is clinically acceptable for most head-and-neck, thorax and pelvic OARs, with time saved even after editing. Target volumes, post-operative beds and nodal regions still need substantial human editing.
• Foundation-style models (TotalSegmentator, promptable 3D models) widen coverage but are not CE-marked RT products by themselves.
• Dose-prediction networks approximate the TPS dose well enough to flag sub-optimal plans and warm-start optimisation; open benchmark challenges (OpenKBP) standardised evaluation.
• Online adaptive RT with AI contouring is moving from MR-Linacs to CBCT-based systems; sCT enables MR-only planning.
• Limits: domain shift across scanners, contrast, implants and unusual anatomy; silent quality drift after protocol changes; overreliance by junior staff; contour metrics (Dice) correlate poorly with dose impact.

Shared resources & combinations

• Same segmentation stack and DICOM plumbing as HLTH-001 (radiology), HLTH-002 (pathology tiling), HLTH-023 (reconstruction/sCT); share a GPU server (HW3), PACS/DICOM gateway and de-identification pipeline.
• DATA-005 (synthetic data) and GEN-020 help augment rare anatomy; HLTH-004 for protocol and guideline retrieval.
• Imaging, TPS, oncology information system (ARIA, Mosaiq) and QA tools form one integrated workflow.

Effort, cost & KPIs

• Path A: 2-6 months: commissioning, validation on local patients, workflow redesign, training; licence per site or per patient plan.
• Path C/D: 6-18 months plus regulatory work; the dominant cost is clinician time for curation and validation.
• KPIs: contour edit time, Dice/surface-Dice vs. clinical contours, clinically relevant dose difference, plan-generation time, plan-quality score variance, proportion of plans passing first review, adaptive fractions delivered, near-miss incidents.

People & change

Involve management and the head of radiation oncology from the start, because the budget and clinical responsibility sit there. Radiation oncologists, medical physicists and radiographers are the real users and must co-design the validation and the approval workflow; bring in IT and IT security early for PACS/TPS integration, and the data protection officer (DPO) before any data leaves the clinic.

Works council: depends. A TPS with auto-contouring logs who edited or approved which plan, which under §87(1) no. 6 BetrVG is enough for co-determination, because the Federal Labour Court only requires objective suitability for monitoring. §90(1) no. 3 requires timely information when work processes change with AI. In public hospitals the staff council (Personalrat) applies. Consider an expert under §80(3).

Data protection: patient data are special-category data (Art. 9 GDPR); a DPIA is likely, plus an AVV (Art. 28 GDPR) with every vendor that accesses data, for example for remote support.

AI literacy: the AI Act Art. 4 obliges deployers to take measures that support staff AI literacy; train on typical errors (missed small structures, anatomy variants) and on automation bias. Communicate that the AI proposes and the physician approves every contour and plan. Start with one tumour site (for example prostate or breast) as pilot group and extend after review.

Guides: GUIDE-004, GUIDE-005, GUIDE-006.

Risks, regulation & compliance

• EU AI Act: high-risk (Art. 6(1), Annex I) because the software is a medical device under MDR (typically class IIa/IIb) that requires notified-body conformity assessment. Obligations apply on the staggered AI Act timetable, with Annex I products later than Annex III; check the current Digital Omnibus timeline. Deployers must ensure human oversight and AI literacy (Art. 4, Art. 26).
• MDR (EU 2017/745): clinical evaluation, post-market surveillance, cybersecurity; the intended purpose fixes whether contours are "proposals" or a basis for treatment. Manufacturer changes (retraining) are significant changes needing predetermined change control.
• GDPR: special-category health data (Art. 9); DPIA, processing agreements, EU hosting, pseudonymised DICOM for research.
• Clinical governance: radiation protection law (e.g. StrlSchG in Germany), medical-physics expert role, independent plan checks, local commissioning and periodic QA of the AI model per ESTRO/AAPM guidance.
• Mitigations: mandatory contour review, automated sanity checks (laterality, missing organs), drift monitoring.

Getting started

1. Define scope: sites (e.g. head and neck, prostate, breast) and which steps (contouring, dose prediction, planning).
2. Shortlist 2-3 CE-marked products compatible with your TPS and oncology information system.
3. Commission on 20-50 recent local cases; compare edit time and dose impact against clinical contours.
4. Write the SOP: who reviews, which checks are automatic, how errors are reported; train staff.
5. Monitor KPIs quarterly and re-validate after any software or protocol change.

Sources

1. ESTRO - European Society for Radiotherapy and Oncology - https://www.estro.org/
2. Regulation (EU) 2017/745 (MDR) - https://eur-lex.europa.eu/eli/reg/2017/745/oj
3. MVision AI - https://www.mvision.ai/
4. Therapanacea ART-Plan - https://www.therapanacea.eu/
5. Varian Ethos adaptive therapy - https://www.varian.com/products/adaptive-therapy/ethos
6. RaySearch RayStation - https://www.raysearchlabs.com/raystation/
7. nnU-Net - https://github.com/MIC-DKFZ/nnUNet
8. TotalSegmentator - https://github.com/wasserth/TotalSegmentator
9. Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj