Reliability and comparative accuracy of AI-supported muscle segmentations by medical imaging and radiation therapy students.

Introduction Quantifying muscle tissue through segmentation provides insight into patient health and supports personalized care. However, muscle segmentation is labor-intensive and expert-dependent, while fully automated tools may not be widely available. This study evaluates the reliability and accuracy of AI-assisted muscle segmentation by medical imaging and radiation therapy (MIRT) students. Methods Dutch MIRT BSc students used the 3DSlicer extension TomoSAM to segment muscle tissue. Their results were compared to expert radiologist segmentations using the Dice Similarity Coefficient (DSC). The effect of user experience on segmentation performance was also evaluated. Results The 1678 unique muscle groups segmented by 42 participants resulted in DSC scores with moderate to high overlap between student and expert segmentations (median 0.734-0.908). User experience had minimal impact overall but was significant for the erector spinae and rectus abdominis. Common errors included inclusion of non-muscle anatomy, often due to false positives from TomoSAM. Conclusion TomoSAM-based segmentations showed promising reliability and comparative accuracy. User experience had a limited effect on segmentation accuracy, and certain muscle groups showed errors due to tool limitations. The reduction in workload with TomoSAM is limited due to false positives and sensitivity to limited contrast between structures. Practical implications TomoSAM can reduce segmentation time, but requires preprocessing and manual correction, especially in low-contrast areas. Ensuring quality control and user accountability is essential. The study highlights the need to balance automation with human oversight, especially in different clinical scenarios.