Main Session
Sep
28
SS 25 - From Assessing Gaps to Treatment Delivery: Scaling Safe Radiotherapy in Resource-Limited Settings
Presenter(s)
Baozhou Sun, PhD, MBA - Baylor College of Medicine, Houston, Tx
T. Qin1, S. Sharma2, S. Kibudde3, A. Kavuma3, M. Minjgee4, E. Vanchinbazar4, Y. Han2, S. Liang2, X. Feng5, Y. Ding6, D. A. Hamstra2, and B. Sun2; 1Department of Radiation Oncology, Baylor College of Medicine, Houston, TX, 2Department of Radiation Oncology, Dan L. Duncan Comprehensive Cancer Center, Baylor College of Medicine, Houston, TX, 3Division of Radiation Oncology, Uganda Cancer Institute, Kampala, Uganda, 4National Cancer Center of Mongolia, Ulaanbaatar, Mongolia, 5Carina Medical LLC, Lexington, KY, 6Carina Medical LLC, Ashburn, VA
Purpose/Objective(s):
Cervical cancer remains a high-incidence disease in Africa and other low- and middle-income countries (LMICs). While 3D image-guided adaptive brachytherapy (IGABT) is the gold standard, many clinics in LMICs still rely on 2D brachytherapy planning due to limited local expertise and the significant time required for manual 3D treatment planning. Automated AI-based tools have the potential to bridge this gap by streamlining the 3D workflow. In this study, we evaluated the performance and efficiency of an AI-based tool designed for automated 3D brachytherapy planning in LMIC clinical settings. Our hypothesis is that implementing AI can significantly streamline workflows and improve dosimetric quality for brachytherapy plans in LMICs.Materials/Methods:
Fourteen CT-based tandem and ovoid (T&O) brachytherapy datasets were retrospectively collected from clinics in Uganda and Mongolia. All plans were prescribed 24 Gy in 3 fractions to the High-Risk Clinical Target Volume (HR-CTV). An AI framework was employed to perform three primary tasks: (1) automated segmentation of the HR-CTV and Organs at Risk (OARs: bladder, rectum, sigmoid, and bowel); (2) automated applicator digitization; and (3) automated dose optimization. The AI models were trained on expert-labeled datasets from US-based institutions and tested on LMIC datasets to evaluate performance in resource-limited settings. AI-generated plans were compared against original manual clinical plans using the Dice Similarity Coefficient (DSC) for geometric accuracy and dosimetric metrics (D90 for HR-CTV; D2cc for OARs). Total end-to-end planning time was also compared.Results:
The AI-based tool successfully generated 3D plans for all 14 cases. Mean DSC values for AI-automated contours were 0.69 ± 0.07 (HR-CTV), 0.70 ± 0.07 (rectum), and 0.72 ± 0.16 (bladder). Geometric comparison for the sigmoid and bowel was precluded by the frequent merging of these structures in the original LMIC clinical plans. Dosimetrically, the AI plans demonstrated improved target coverage, with a mean HR-CTV D90 of 8.35 ± 1.02 Gy, compared with 7.47 ± 0.88 Gy in the manual plans. Furthermore, AI plans achieved superior OAR sparing across all structures; mean D2cc values for AI vs. manual plans were rectum (2.41 ± 1.03 Gy vs. 3.26 ± 1.29 Gy), bladder (4.51 ± 2.08 Gy vs. 5.82 ± 1.59 Gy), sigmoid (2.53 ± 0.73 Gy vs. 2.72 ± 1.32 Gy) and bowel (1.23 ± 0.74 vs. 3.82 ± 1.45 Gy). The average end-to-end AI planning time was 9 minutes, about 90% reduction from the manual average of 95 minutes.Conclusion:
An AI-based auto-planning tool trained on high-resource data demonstrated high accuracy and robust generalizability when applied to LMIC clinical cases. The tool not only significantly reduced planning time but also improved dosimetric quality over manual planning. This technology offers a scalable solution to "leapfrog" from 2D to 3D brachytherapy in resource-limited settings, potentially expanding global access to high-quality, life-saving radiotherapy.