AI could reduce radiotherapy workload in low-resource settings, but implementation evidence remains limited

A critical review of 11 studies found major workflow gains from AI-assisted contouring and planning, but limited evidence of sustainable implementation in resource-constrained settings.

KEY POINTS

  • This critical review evaluated artificial intelligence applications in radiation oncology across low- and middle-income countries (LMICs), searching PubMed, Embase and Cochrane Library for publications from 2010 through April 2026. From 488 records, 11 publications were included: five original research studies, four reviews, one survey and one study protocol. The evidence focused predominantly on automated contouring, treatment planning and workforce efficiency, with limited directly measured clinical outcomes.
  • AI-assisted auto-segmentation produced substantial time savings in the available LMIC studies. Experience from Uganda and Mongolia demonstrated reductions in organ-at-risk contouring time from approximately 57–84 minutes to about two minutes per patient, with generally comparable contour quality. The review reported estimated annual workload savings of approximately 1,000 person-hours in one implementation, although workload estimates differed between cited studies and did not necessarily include downstream physician review.
  • Automated treatment planning also reduced technical workload. At the Uganda Cancer Institute, RapidPlan-based knowledge-guided VMAT planning across six disease sites reduced mean planning time from 43.2 to 29.5 minutes, an approximately 33% reduction (p<0.01), while maintaining acceptable target coverage and OAR dosimetry across most evaluated sites. Separately, Radiation Planning Assistant evaluations reported high plan acceptability, including 96% for head and neck VMAT plans and approximately 99–100% for cervical VMAT plans with no or minor modifications, although these assessments were not exclusively LMIC-based.
  • Local clinical evaluation remains essential because algorithmic performance may not transfer between healthcare settings. The randomized ELAISA educational study, conducted across 23 LMIC radiotherapy institutions involving 97 radiation oncologists, evaluated AI-assisted contouring and training, while the ARCHERY study was designed to enroll 990 patients to examine planning acceptability and economic implications. The latter was represented by a study protocol in the review, so its results cannot be interpreted as demonstrated implementation success.
  • Infrastructure and workforce limitations extend beyond the availability of AI software. The authors identify unreliable electricity and internet connectivity, outdated CT equipment, fragmented electronic medical records, limited PACS infrastructure, insufficient locally annotated datasets, and licensing costs as major barriers. Commercial AI modules may require substantial recurring expenditure, while high-income-country training datasets may not reflect local scanner technology, disease presentation or clinical workflows.
  • Evidence of improved clinical outcomes remains scarce. Although auto-contouring and planning tools have reduced treatment preparation time and improved workflow consistency, the reviewed literature did not establish that AI adoption independently improves tumor control, survival, or treatment-related toxicity compared with conventional workflows. Moreover, economic benefits and increases in patient throughput have not been prospectively established across representative LMIC health systems.
  • The review recommends locally validated, affordable, interoperable and potentially offline-capable AI systems, accompanied by workforce education and explicit human oversight. International partnerships, federated learning and local data governance are proposed as possible implementation pathways. However, conclusions are limited by heterogeneous study designs, inclusion of secondary reviews, sparse original LMIC-specific evidence, and extrapolation from high-income settings.

CLINICAL TAKEAWAY

AI-assisted contouring and planning could meaningfully reduce workload in resource-constrained radiotherapy departments, where specialist time is a major limitation to treatment delivery. However, deployment should prioritize local validation, affordable infrastructure, training and clinical oversight, while prospective research must establish whether improvements in efficiency translate into better patient access and outcomes.

SOURCE

International Journal of Radiation Oncology, Biology, Physics

Browse more research Suggest a correction