Is AI ready for cervical-radiotherapy workflows?
A review of 52 studies finds encouraging automation for contouring, planning and applicator reconstruction. Most evidence remains retrospective and single-centre, with high bias concerns and little independent validation or proof of patient benefit.
The short version
- What happened
- The PRISMA review searched PubMed, Scopus and Web of Science for 2010–2026 evidence, screened 353 records and included 52 studies after full-text review.
- Who it affects
- Patients may eventually spend less time waiting for planning, but current studies do not show better survival, fewer complications or safe autonomous treatment.
- What remains unproven
- Large methodological and clinical heterogeneity prevented direct model comparison or meta-analysis
Keep in mind
Large methodological and clinical heterogeneity prevented direct model comparison or meta-analysis
Many studies were retrospective, small and single-centre
Independent external and prospective validation was uncommon
Technical overlap and dose metrics do not establish patient benefit
Publication bias and inconsistent reporting may overstate maturity
Source analysis · 1 linked source. Primary analysed source date: 15 October 2026.
Read the full report · 6 min The evidence, methods, implications and limitations
The answer: technically promising, clinically immature
AI can already automate parts of image-guided cervical interventional radiotherapy with impressive results in selected datasets. Across the review, segmentation systems sometimes reached Dice similarity coefficients near or above 0.90 for particular organs at risk or target volumes. Individual studies also reported faster processing, less contouring variation, accurate reconstruction of applicators and useful support for applicator choice. These are meaningful engineering signals in a complex, time-sensitive treatment workflow.
They do not show that an autonomous system is ready to plan or deliver treatment. The 52 studies differed in imaging, anatomy, clinical practice, model design and outcomes, so their headline scores cannot be ranked as a common league table. Most tools were retrospective proofs of concept. Independent multicentre validation and prospective clinical evaluation were rare or absent, and the review repeatedly found high risk of bias in data selection and analysis. A high overlap score is not the same as fewer complications, better local control or longer survival.[1]
How the review assembled the evidence
The authors used PRISMA 2020 and defined the method before searching PubMed, Scopus and Web of Science for original studies published from 2010 through 2026. Eligible work had to evaluate AI in image-guided cervical interventional radiotherapy, the modern form of cervical brachytherapy, including segmentation of targets and organs at risk, treatment planning, applicator reconstruction, image registration or applicator selection. The search found 353 records, removed 127 duplicates and screened 226 titles and abstracts. Sixty-eight full texts were assessed and 52 studies were included.
Quality was evaluated with tools matched to the task. Segmentation, reconstruction and registration studies were assessed as diagnostic-style evidence, while planning and applicator-selection studies were assessed with the 2025 PROBAST+AI framework. That matters because the review did more than collect the best reported scores: it asked whether participants, predictors, reference standards and analyses would support transport into practice. The result is a useful map of technical progress alongside the design weaknesses that limit deployment.[1]
Where AI appears most useful
Segmentation is the most visible application. Clinicians must delineate the tumour and nearby bladder, rectum, bowel and other structures on imaging before calculating dose. Models that generate a reliable first contour could reduce repetitive work and make planning more consistent, provided a qualified clinician reviews every boundary. The best scores were not uniform: structures varied in shape, contrast and imaging protocol, and the review warns that apparently similar Dice values may conceal clinically important errors in different locations.
Planning and reconstruction target different bottlenecks. Planning tools try to predict or optimise a dose distribution, while reconstruction systems identify the implanted applicator and its geometry on scans. Applicator-selection models aim to support a choice before treatment, and registration models align images across time or modalities. Each could shorten work, but each also creates a different failure mode. An inaccurate contour, reconstructed channel or registered image can propagate through the dose calculation, so quality assurance must be task-specific rather than based on a generic AI accuracy label.[1]
Bias concerns are concentrated where deployment matters
For planning studies, 11 of 14 were judged at intermediate risk in the participants-and-data domain, largely because datasets were small, retrospective and single-centre. The analysis domain was the larger concern: limited samples, potentially inadequate independent evaluation, overfitting or leakage and missing external validation. Reconstruction studies were judged predominantly high risk overall, and none had independent external validation. All three image-registration studies had high risk in the index-test domain and no independent external validation.
Applicator-selection studies had the same structural problem. Predictors were often objectively defined and outcomes could be clinically grounded, yet small local datasets and insufficiently independent evaluation left uncertain generalisability. These findings explain why a polished demonstration is not enough. A model can learn scanner settings, annotation habits, applicator brands or patient mix from its development centre. It needs to work when those conditions change and when clinicians encounter difficult cases that were uncommon in the training set.[1]
What this means for patients and clinical teams
For patients, the near-term benefit is more likely to be shorter planning time and more consistent preparation than a fully automated treatment. Any service introducing AI should preserve clinician responsibility, display the generated contour or plan for review and measure corrected errors. Patients should not be told that technical similarity proves a safer or more effective treatment. The review did not establish reductions in toxicity, recurrence, waiting time or inequality of access.
For radiation oncologists, radiographers, physicists and dosimetrists, the central question is where the tool sits in the workflow. Evaluation should include time saved after corrections, disagreement by structure, failure detection, override rates and performance on unusual anatomy. Procurement should require site-specific acceptance testing and monitoring after software or scanner changes. A model that saves ten minutes on routine cases but creates rare undetected dose errors may not be a net improvement.[1]
Evidence limits and what would change the assessment
The review itself is limited by the underlying studies and by publication in an early formatted state. Heterogeneity prevented pooled performance estimates, and favourable technical studies may be easier to publish than unsuccessful replications. Study-level sample sizes and clinical endpoints were inconsistent, so the review cannot identify one model or threshold as ready for universal adoption. The authors reported no financial support or commercial conflict, although one author served on the journal's editorial board and the publisher states that this did not affect review or decision-making.
Confidence would rise with locked models tested prospectively across hospitals, scanners, applicator types and patient populations, using predefined failure rules and independent annotations. Trials should compare AI-assisted and standard workflows on corrected planning time, clinically important contour errors, dose measures, toxicity, local control and staff workload. Until those results exist, the evidence supports supervised workflow research and careful local evaluation—not autonomous treatment planning or claims of improved cancer outcomes.[1]