How far has artificial intelligence been deployed in low-resource radiation oncology?

A systematic review finds LMIC radiation oncology AI stays at task-level help, with no monitored autonomous deployment.

Direct answer

In low- and middle-income countries, AI in radiation oncology has mostly reached only task-level assistance: of 18 included studies, nine (50.0%) were Level 1 and six (33.3%) were Level 2 clinician-supervised workflow integration, while just one (5.6%) reached cross-stage coordination and none met the criteria for monitored, governed autonomous operation. Validation was predominantly retrospective, and no study reported continuous post-deployment performance monitoring.

1sources cited

This article was generated with WisPaper-powered search and paper analysis.

A five-level spectrum of workflow integration

The review asked a narrower question than whether AI can contour a tumor or generate a plan. It asked how far each reported AI system was actually embedded in the radiotherapy workflow in low- and middle-income countries (LMICs), and whether the surrounding validation and governance would support safe scale-up.

To answer that, the authors built a radiation oncology-specific integration spectrum from Level 0 to Level 4. Level 0 means no AI output is used in clinical or clinical-simulation workflows, including studies that only surveyed perceptions or readiness. Level 1 is isolated task assistance, such as contouring without downstream coordination. Level 2 embeds approved outputs into a fixed workflow sequence with mandatory human review. Level 3 adds predefined rules or conditional logic so intermediate outputs determine later actions, with structured human checkpoints. Level 4 requires autonomous multi-stage operation plus continuous performance monitoring, drift detection, escalation thresholds, audit logs and formal governance.

The spectrum is deliberately workflow-centric rather than a measure of algorithmic sophistication. Assignment depended on demonstrated operational behavior, not stated intent, and two reviewers classified studies independently, agreeing on 16 of 18 (88.9%) before resolving the two discrepancies by consensus. That design choice matters: a highly accurate model used as a standalone tool is still Level 1 under this scheme.

Eighteen studies from 1,309 records

Following PRISMA 2020 guidelines, the authors searched six databases plus grey literature such as IAEA and WHO technical reports and conference proceedings, covering January 2000 through December 2025, with the final search run on January 16, 2026. Searches yielded 1,309 records; after removing 113 duplicates, 1,196 were screened by title and abstract and 1,156 were excluded. Thirty-nine full-text reports were assessed, and 21 were excluded, most often because they were not focused on LMIC settings (n = 8), were book chapters (n = 5), or were published in a language outside the eligibility criteria (n = 5).

Eighteen studies entered the final review, all peer-reviewed; no grey-literature record met the inclusion criteria. Because several studies covered multiple countries or institutions, the authors also counted implementation instances, defined as each country where a system was deployed or evaluated. That produced 25 instances in Latin America, 12 in Africa, 5 in Asia and one mixed-region deployment, a total exceeding the 18 studies. The authors treated these regional counts as descriptive geography only, not as an inferential comparison, because implementation instances are not statistically independent.

Half of the studies stop at isolated tasks

The distribution across integration levels is the paper's central result. Two studies (11.1%; 95% CI 3.1%–32.8%) were Level 0, covering readiness, governance, professional perceptions and adoption barriers without operational AI use. Nine (50.0%; 95% CI 29.0%–71.0%) were Level 1, mostly contouring or segmentation without downstream workflow coordination. Six (33.3%; 95% CI 16.3%–56.3%) reached Level 2, where approved outputs feed subsequent predefined activities. One (5.6%; 95% CI 1.0%–25.8%) reached Level 3, and no study (0%; 95% CI 0%–17.6%) met Level 4.

The single Level 3 case was the Radiation Planning Assistant, which coordinates sequential contouring, planning and quality-assurance activities with structured human checkpoints. Application domains followed the same pattern: AI tools concentrated on contouring and segmentation and on treatment planning, with fewer studies addressing quality assurance or readiness, and most studies covering a single workflow component. The authors describe this as a fragmented deployment landscape rather than a broad transformation of radiotherapy practice.

Cloud delivery is common, prospective validation is not

Deployment models split unevenly. Twelve studies (66.7%) delivered AI through web-based or cloud-enabled platforms, often to reduce local hardware requirements and reach institutions with limited computational infrastructure; six (33.3%) described locally deployed or offline systems, typically where institutions had more control over hardware and data storage.

Validation was the weaker dimension. Ten studies (55.6%) validated on datasets generated from LMIC clinical settings, while eight (44.4%) relied on external or mixed datasets. In many cases validation was retrospective and centered on geometric accuracy metrics such as Dice similarity coefficients rather than prospective clinical performance, and only a few studies described real-world or prospective validation in routine workflows. Reporting on validation and monitoring was highly inconsistent across the dataset.

Critical appraisal found generally adequate methodological quality for the technical validation and implementation studies, but most were retrospective and several used small or single-institution cohorts. The authors flagged potential selection bias in one perception survey that used convenience and snowball sampling without a defined sampling frame or calculable response rate. These are limits on what the included evidence can show, not on the review's search process.

Workflow coupling and governance limit scale-up

The authors argue the pattern reflects system-level constraints rather than a lack of innovation. Contouring, segmentation and planning are comparatively modular, can reduce workload at specific stages, and can be evaluated independently, which makes them pragmatic first targets in resource-constrained settings. But radiotherapy is sequential and interdependent: simulation, contouring, planning, quality assurance and delivery depend on reliable handoffs and on the integrity of upstream information.

That structure raises the cost of deeper integration. Heterogeneous information systems, manual handoffs, intermittent connectivity and limited technical support make cross-stage coordination harder to implement and maintain, and as integration increases, failures can propagate further downstream before anyone detects them. The absence of clearly defined escalation logic and workflow-level validation helps explain why most current implementations remain deliberately bounded and clinician-supervised.

Governance is a parallel constraint. Formal mechanisms for accountability, approval of model updates, auditability or oversight of increasingly autonomous behavior were rarely described, which matters more where limited staffing and reduced system redundancy make recovery from an AI-related failure harder. The authors present task reliability, workflow integration and system assurance as coupled rather than sequential requirements.

Absence of Level 4 is not proof of absence

Two boundaries qualify the headline distribution. First, the authors caution that the absence of Level 4 systems should not be read as proof such capabilities do not exist in practice; post-deployment monitoring, governance or lifecycle-management infrastructure may simply be incompletely reported in the included literature. The zero count describes what was documented, not necessarily what was built.

Second, this review is described as the first peer-reviewed application of the proposed integration spectrum. Although criteria were predefined and classification was independent, the framework requires interpretive judgment, and its reproducibility and generalizability have not yet been independently evaluated. The level percentages should therefore be read as a structured description of a heterogeneous literature, not as a validated measurement instrument. The review also performed no inferential statistical testing and no sensitivity analyses, given heterogeneity in study design, application domain, validation approach and outcomes.

Prospective workflow evidence is the missing link

The paper's forward-looking priority follows directly from where the evidence is thin. The authors call for prospective, multi-institutional implementation studies across diverse LMIC settings that evaluate both technical performance and real-world workflow effects, including throughput, editing burden, failure modes and sustainability over time. They also call for greater sharing of representative datasets to support external validation, with federated approaches where direct data sharing is not feasible, and for governance frameworks covering accountability, validation, escalation, model updates, auditing and post-deployment monitoring.

The distinguishing judgment is that in LMIC radiation oncology, scalable AI should be assessed by whether it can be integrated reliably into clinical workflows and sustained safely over time, not by task accuracy alone. On the current evidence, the gap between a working contouring model and a governed multi-stage system is where almost all of the reviewed deployments sit, and closing it depends on validation, monitoring and governance advancing together with the models themselves.

Sources used in this answer

1

Implementing and Scaling Artificial Intelligence in Low-Resourced Radiation Oncology: A Systematic Review of Deployments

Of 18 included studies on AI in LMIC radiation oncology, 9 (50.0%) were Level 1 task-level applications, 6 (33.3%) were Level 2 clinician-supervised workflow integration, 1 (5.6%) reached Level 3 cross-stage coordination, and none met Level 4 monitored autonomous operation; validation was predominantly retrospective and continuous performance monitoring was not reported.