AI prepared radiotherapy plans that passed expert review in an international cancer study

AI-generated radiotherapy plans met a high clinical standard in more than 95% of cervical-cancer cases and 85% of prostate-cancer cases in the ARCHERY study, according to results presented at ESTRO in May 2026. The software’s achievement was completing detailed preparation that ordinarily requires substantial specialist time.[1]

The international study recruited 1,029 adults at seven hospitals in India, Jordan, Malaysia and South Africa. There were 331 participants with cervical cancer, 351 with prostate cancer and 347 with head-and-neck cancer. The reported cervical and prostate findings assessed the quality of the proposed plans; participants received their hospital’s usual treatment plans.[2]

The work before the first treatment

Before radiotherapy begins, a team must identify the tissue to treat and the healthy structures to protect. It then designs the position, shape and intensity of the radiation beams. The planning scan is therefore the beginning of a sequence of decisions about anatomy and dose, rather than a picture sent directly to a treatment machine.[3]

The Radiation Planning Assistant, developed at MD Anderson Cancer Center, combines two forms of AI. Deep-learning models identify anatomical structures on CT images. Knowledge-based planning models draw on previously prepared plans to guide a new dose distribution. A July 2026 technical report describes training data from more than 1,000 prostate plans and 110 cervical-cancer plans.[4]

MD Anderson’s account of its 2024 launch describes clinicians submitting a deidentified CT scan and a service request. It also records initial clinical use in South Africa. This gives the software a practical delivery route: a hospital can use remote planning capability and review the returned work locally.[5]

Testing whether the complete plan is acceptable

ARCHERY evaluated the automated pathway using the same CT scan as the hospital’s standard pathway. Two independent reviewers, blinded to the hospital’s identity, assessed the AI-generated contours and plans against an international protocol. A third reviewer resolved disagreements. Recruitment ran from January 2024 to January 2026, giving the study a prospectively assembled, multicountry population.[2]

The protocol required acceptable contouring and acceptable dose planning together. A plan could pass with minor edits that were not expected to have a major clinical effect. Errors requiring major correction counted as unacceptable. This is an assessment of whether the combined work is clinically usable, with a threshold more demanding than simply producing plausible-looking outlines.[3]

That definition also explains the outcome being reported. ARCHERY did not measure whether using the AI plan improved survival or reduced treatment complications. Its published protocol explicitly kept participants on standard plans. The enrolled group sizes establish the study’s scale; the May announcement did not provide exact analysed counts behind the two reported acceptability percentages.[3][2][1]

The software also has a smaller, separate validation

The July technical report evaluated 50 test cases: 40 prostate and ten cervical. Radiation oncologists retrospectively assessed the results. For prostate cancer, 70% of target outlines and 73% of treatment plans were acceptable without edits; the corresponding cervical figures were both 80%. These are a separate development evaluation, with a stricter no-edit description, and should not be pooled with ARCHERY’s results.[4]

The distinction is useful when assessing automation. A model can complete most of a demanding job while still leaving worthwhile corrections for an expert. Measuring the quality of the whole plan, and the changes it needs, shows where assistance fits into clinical work.

Making specialist time go further

UCL’s account quotes the trial’s chief investigator describing automated planning in just over an hour. Conventional work may occupy many staff hours spread across a longer calendar period because specialists are handling other patients. The opportunity is to reduce that preparation burden, although the announcement does not establish a uniform reduction in patient waiting time.[1]

The current RPA platform describes its purpose as expanding planning access in settings with limited staff and resources. Its stated aim is to offer the service without charge to clinics that cannot access commercial alternatives. Bringing several planning stages together is central to that ambition: the useful output is a treatment proposal the local team can examine and complete.[6]

Report an error or suggest a correction