Using AI to Improve Radiation Therapy Planning

Vollständiger Abstract

Worum geht es in dieser Arbeit?

In radiation therapy, clinicians must outline nearby organs, known as organs at risk (OARs), to limit unwanted radiation exposure. This process is traditionally done by hand and requires significant time and resources. Artificial intelligence (AI) tools can automatically generate these outlines, potentially improving efficiency. However, their clinical reliability across different treatment sites must be carefully evaluated. We analyzed various patient cases across multiple anatomical sites, including thoracic, head and neck, breast, and pelvic regions. Clinician-created manual contours were compared with AI-generated contours from two treatment planning systems: Ethos-2 and RayStation 2023B. Quantitative agreement was assessed using standard image comparison metrics, along with an evaluation of radiation dose differences to critical organs. In addition, a qualitative clinical evaluation was performed for lung cases, where radiation oncologists evaluated contour acceptability and clinical use. Across all treatment sites, AI contours showed strong agreement with manual contours and resulted in minimal differences in delivered radiation dose. Larger and commonly contoured organs, such as the lungs and heart, demonstrated particularly good performance. Some smaller and complex structures, especially in head and neck and GI cases, showed reduced agreement, though these differences generally had little impact on dose. In the lung qualitative review, clinicians found most AI contours acceptable, but certain structures (notably the esophagus and heart) were more frequently flagged due to anatomical inaccuracies rather than safety concerns. In conclusion, AI-based segmentation performs well across multiple anatomical sites and is dosimetrically safe for clinical use. However, clinician review remains essential, as quantitative accuracy alone does not fully capture clinical acceptability. Continued model refinement will further enhance the clinical integration of AI contouring tools.