After a structured clinical-alignment process, AI automation became the default pathway for prostate plans. Automation volume more than doubled in the first seven months.
Automation adoption in radiotherapy planning is usually framed as a plan-quality problem. Our experience at a US community cancer center suggests it is an alignment problem — and that a short, data-driven alignment loop can solve it.
Presented at ASTRO 2026 (Boston, September 26–30). Yunes, Napier, Garg, Dong, Kaufman, Mc Loone, Price, Schuler & Fox — Mass General Brigham Cancer Institute, UMass Chan Medical School – Baystate, and Lumonus.
The barrier isn't plan quality
A decade of studies has shown that automated radiotherapy planning reliably improves dosimetric quality, planning consistency, and workflow efficiency. Yet scaled adoption has stalled across the field. Our hypothesis was that the key barrier isn't plan quality, but unmeasured variability in physician planning preferences: radiation oncologists differ — between institutions, and within the same department — in how they judge a plan and where they draw trade-offs. Until those preferences are made visible and built into the automation, every automated plan is judged against a standard nobody has written down.
To test this, we ran a structured clinical alignment process at Baystate Medical Center before deploying Lumonus AI Dosimetry clinically, and tracked what happened to adoption afterwards.
Methods — the clinical alignment loop
- Select 20 randomly selected prostate plans treated in the prior year, spanning conventional prostate, prostate + nodes, and prostate SABR.
- Automate AI plans generated with no manual edits — automation tested in isolation, deliberately without a human in the loop.
- Benchmark Population DVH analysis and consensus planning metrics (NRG GU005; NRG Pelvic Lymph Node Consensus) against the previously approved clinical plans.
- Review A quality report reviewed with physicians, physics, and dosimetry — where implicit preferences surface and get discussed.
- Tune + repeat Automation blueprints refined to the institution's stated preferences, then re-benchmarked until aligned — then deployed.
- The loop is iterative: review, tune, re-benchmark until aligned.
What alignment surfaced
- Hidden normalization split 4 of 10 prostate and 2 of 5 SABR plans were normalized to D98, the rest to D95 — unknown until measured. Models were aligned to the institutional choice, D95.
- Quality held At cohort level, automated plans at least matched clinical plans on target coverage and organ-at-risk sparing.
- Consistency gained Improved rectal and/or bladder sparing, with narrower interquartile ranges across the cohort — the same plan quality, more predictably.
- Trade-offs made explicit The review sparked debate on relative OAR priorities — notably bladder versus rectum dose in prostate + nodes cases — turning implicit individual preferences into explicit institutional decisions.
What happened after deployment
Alignment concluded in December 2025 and the tuned models went live in January 2026. Two numbers tell the adoption story:
Monthly automated prostate plan volume more than doubled in the first seven months, and the share of plans needing no dosimetrist edits rose from 13% to 47%.
Data: January–July 2026, single US community RT center. Persistent major edits are expected and managed by the dosimetrist-in-the-loop workflow — automation ≠ autonomy. Every plan is still reviewed and, where required, refined by a certified medical dosimetrist before it reaches a physician.
Explore the data behind the results (click on an option)
Take it to your clinic
Nothing in this process is specific to one vendor's model or one treatment planning system. The recipe is simple and portable: benchmark automation against your own approved plans, put the comparison in front of the people whose preferences actually decide plan acceptance, tune to what they tell you, and repeat until aligned. In our experience, that short loop converts physician skepticism into measured preferences — and preferences into adoption.
If you'd like the full clinical alignment report behind this work, or want to talk about running clinical alignment at your center, get in touch.
Reference: Yunes MJ, Napier T, Garg S, Dong Y, Kaufman S, Mc Loone P, Price M, Schuler T, Fox T. A Systematic Data-Driven Process to Support Deployment of Radiation Therapy Planning Automation. ePoster, ASTRO 68th Annual Meeting, Boston, 2026. Consensus metrics: NRG GU005; Hall et al., IJROBP 2021.
Disclosures: P. Mc Loone, M. Price, T. Schuler and T. Fox are employees of Lumonus. Adoption data reflect January–July 2026 at a single US community RT center; dosimetric comparisons reflect the 20-plan clinical alignment cohort described above.
About the interactive figures: Medians and interquartile ranges are as published in the ASTRO 2026 abstract. Box plots (type-7 quartiles, 1.5×IQR whiskers) and population DVH curves are computed from the per-plan data of the clinical alignment cohort; DVH curves are adaptively downsampled for display with a maximum deviation below 0.25% volume.
