Investigation on performance of multiple AI-based auto-contouring systems in organs at risks (OARs) delineation

Area of Science:

• Medical Physics
• Radiotherapy
• Artificial Intelligence

Background:

• Manual organ at risk (OAR) contouring is labor-intensive and prone to inter-observer variability.
• AI-based auto-contouring offers a potential solution to improve efficiency and consistency in radiotherapy planning.

Purpose of the Study:

• To evaluate and compare the performance of multiple AI-based auto-contouring systems for OAR segmentation across diverse anatomical sites.
• To assess the clinical acceptability and potential utility of AI-generated contours in radiotherapy.

Main Methods:

• Seven AI segmentation systems were tested on 42 clinical cases.
• Performance was evaluated using Dice similarity coefficients and Hausdorff distance against expert manual contours.
• Analyses covered various anatomical sites including head and neck, brain, lung, breast, pelvis, and abdomen.

Main Results:

• Radiotherapy AI demonstrated superior performance in head and neck and brain OARs.
• No single AI system showed overall superiority across all anatomical sites.
• AI systems achieved contours comparable to expert delineations, with reduced performance on small/complex structures.

Conclusions:

• AI auto-contouring systems can produce clinically acceptable OAR contours, potentially streamlining radiotherapy workflows.
• Clinical implementation requires careful review of AI-generated contours, especially for intricate anatomical regions.
• The study provides a framework for comparing AI contouring software and for ongoing quality assurance.