Impact of AI-Based Auto-Segmentation on Radiotherapy Processes in Low and Middle-Income Countries

Abstract

Radiotherapy processes require significant human resources and expertise, creating a barrier for rapid deployment in low and middle-income countries (LMICs). Optimal radiotherapy (RT) relies on accurate segmentation of tumor targets and organs-at-risk (OARs) during the RT planning process. This study reports the impact of AI-based auto-segmentation on RT processes in an LMIC. Ten patients including five head and neck (HN), and five prostate cancer patients were randomly selected. Their planning CT images were subjected to auto- segmentation using an FDA-approved AI software tool, and manual segmentation by an experienced radiation oncologist from a Sub-Saharan African RT clinic. The control data consisted of contours from an experienced radiation oncologist and dosimetrists at a large academic institution in the US. For prostate cases, the contours included the prostate, seminal vesicles, bladder, rectum, penile bulb, and both femoral heads. For HN cases, the contours included the brain, brainstem, bilateral eyes, lens, optic nerves, cochlea, parotids, optic chiasm, spinal cord, oral cavity, and mandible. The time to complete the segmentation was recorded for both auto-segmentation and manual contours from the LMIC. The DICE similarity coefficients were used for comparative evaluation. The average time for contouring per patient was 2 minutes for AI compared to 57 minutes for manual contouring in the LMIC. When comparing the control data, AI pelvic contours provide a slightly better agreement than LMIC manual contours for all the OARs, with the following mean DICE coefficients for AI vs LMIC manual contours: bladder (0.971 vs 0.958), left femoral head (0.960 vs 0.949), right femoral head (0.959 vs 0.941), rectum (0.880 vs 0.867), prostate (0.836 vs 0.824), seminal vesicles (0.696 vs 0.580), and penile bulb (0.536 vs 0.528). For HN contours, AI provide a better agreement for 7 of 11 OARs than the LMIC manual contours, with the following mean DICE coefficients: brain (0.972 vs 0.982), mandible (0.877 vs 0.925), right parotid (0.847 vs 0.800), left parotid (0.798 vs 0.792), spinal cord (0.837 vs 0.821), left eye (0.875 vs 0.832), right eye (0.867 vs 0.836), brainstem (0.866 vs 852), oral cavity (0.796 vs 0.787), left lens (0.650 vs 0.729) and right lens (0.671 vs 0.682). Neither AI contours nor LMIC manual contours had good agreement with the control data (<0.600) for optic nerves, chiasm, and cochlea due to their small volumes. AI-based auto-segmentation tools are capable of producing contours of comparable quality to those generated by manual segmentation for both pelvic and HN cancer patients in LMICs, while also resulting in substantial time savings. AI-based auto-segmentation holds tremendous potential for improving radiotherapy care in LMICs with limited sources.