236 - AI in LMIC Radiotherapy: Clinical Performance and Implementation Feasibility

Purpose/Objective(s):

Radiotherapy services in low- and middle-income countries (LMICs) face severe workforce shortages and inefficient workflows. In high-income countries (HICs), artificial intelligence (AI) reduces contouring and planning times by 30–65% while improving geometric consistency. In LMICs, where clinicians often manage higher patient volumes, similar efficiencies could meaningfully expand treatment capacity. We evaluated whether AI tools deployed in LMIC settings demonstrate improvements in workflow efficiency, geometric and dosimetric quality, and implementation feasibility.

Materials/Methods:

We conducted a systematic literature review (2000–2025) following PRISMA guidelines. Of 1,196 screened records, 19 met inclusion criteria. Two reviewers independently extracted data in Rayyan with third party adjudication. Risk of bias was assessed using ROBINS-I. Outcomes included workflow domain, disease site, validation approach, deployment model, and clinical impact, and implementation barriers.

Results:

AI applications in LMIC radiotherapy primarily addressed treatment planning (13) and auto-segmentation(6). Across studies, AI improved geometric agreement and workflow efficiency. Reported gains include higher Dice similarity coefficients and reduced editing times. A multi-country educational RCT demonstrated improved head-and-neck OAR performance and faster contouring, with 24% of AI contours accepted without edits and sustained benefits at 6 months. Across studies, FDA-cleared and locally trained auto-segmentation systems achieved clinically acceptable geometric accuracy (DSC ~0.8-0.92), with typically minor remaining edits. Automated planning platforms produced largely acceptable cervical, head-and-neck, and whole-brain plans with substantial reductions in planning time. Deployment models were predominantly cloud- or web-based. Recurrent barriers included bandwidth limitations, inadequate compute capacity, cost constraints, data privacy concerns, and model bias. Enablers included standardized workflows, brief hands-on training, institutional support, and adaptable cloud infrastructure. Cost-effectiveness and validation reporting remained limited.

Conclusion: AI tools in LMIC radiotherapy settings demonstrate consistent improvements in contour quality, plan acceptability, and workflow efficiency, supporting their potential to alleviate workforce shortages and increase treatment capacity. However, inconsistent validation reporting and infrastructural limitations remain barriers. Standardized commissioning, end-to-end testing, and governance frameworks are essential for safe and scalable implementation.

Domain	Key Finding
Autosegmentation	Improved geometric accuracy (DSC ~0.8–0.92) with less editing
Planning	Clinically acceptable plans with substantial time savings
Training (H&N RCT)	24% contours accepted without edits
Implementation	Cloud- or web-based deployment