3793 - Can Open-Source Artificial Intelligence Reduce Disparities In Radiation Oncology? A Pilot Auto-Segmentation Study In a High-Volume Public Center for Female Pelvic Malignancies

Purpose/Objective(s): Contouring is a critical step in radiation therapy and directly influences treatment start times. In high-volume public cancer centers, particularly in low- and middle-income countries (LMICs), workflow inefficiencies may contribute to treatment delays and exacerbate disparities in access to care. Open-source artificial intelligence (AI)-based auto-segmentation tools represent a potentially scalable and cost-effective strategy to optimize contouring workflows. This pilot study evaluated whether AI-assisted auto-segmentation for female pelvic malignancies reduces the time between contouring initiation (CI) and treatment planning (TP), compared with manual contouring.

Materials/Methods: This retrospective pilot study was conducted at a high-volume public cancer center in Lima, Peru. Patients with histologically confirmed female pelvic malignancies (cervical and endometrial cancer) treated with definitive or adjuvant radiation therapy between May and November 2025 were included. Two groups were compared: (a) manual contouring, in which organs at risk (OAR) were delineated by radiation oncologists and residents; and (b) AI-assisted contouring, in which OAR were auto-segmented using open-source artificial intelligence software, while target volumes were contoured manually by radiation oncologists. The primary endpoint was the number of days between contouring initiation (CI) and treatment planning (TP). Continuous variables were reported as mean ± standard deviation and median (interquartile range). Group comparisons were performed using the Mann–Whitney U test.

Results: A total of 62 patients were included, with 31 patients in each group. The mean CI–TP interval was 10.13 ± 5.61 days in the manual contouring group and 9.03 ± 5.46 days in the AI-assisted group. Median CI–TP times were 11 days (IQR: 6–14) and 9 days (IQR: 4–12.5), respectively. The observed approximately 2-day reduction in median CI–TP interval did not reach statistical significance (p = 0.45). Descriptive and visual analyses (boxplot and histogram) demonstrated overlapping distributions between groups, with a trend toward shorter CI–TP intervals in the AI-assisted cohort.

Conclusion: This pilot study demonstrates the feasibility of implementing open-source AI-based auto-segmentation tools for OAR delineation in a high-volume public radiation oncology center in a middle-income country. Although the observed reduction in CI–TP interval did not reach statistical significance, likely due to limited sample size, these preliminary findings generate hypotheses and suggest that open-source AI tools may contribute to workflow optimization. Prospective studies with adequately powered sample sizes are warranted to determine whether this strategy translates into clinically meaningful reductions in treatment initiation times and improved timely access to radiation therapy for female pelvic malignancies in resource-limited settings.