Technical accuracy of AI-based patient autopositioning in Computed Tomography: An evaluation with dose monitoring system data.
Authors
Affiliations (7)
Affiliations (7)
- Medical Physics Unit, USL Toscana Sud-Est, Italy. Electronic address: [email protected].
- Medical Physics Unit, USL Toscana Sud-Est, Italy. Electronic address: [email protected].
- Medical Physics Unit, USL Toscana Sud-Est, Italy. Electronic address: [email protected].
- Medical Physics Unit, USL Toscana Sud-Est, Italy. Electronic address: [email protected].
- Medical Physics Unit, USL Toscana Sud-Est, Italy. Electronic address: [email protected].
- Medical Physics Unit, USL Toscana Sud-Est, Italy. Electronic address: [email protected].
- Medical Physics Unit, USL Toscana Sud-Est, Italy. Electronic address: [email protected].
Abstract
Recent advancements in Artificial Intelligence (AI)-driven algorithms have improved patient alignment in Computed Tomography (CT) imaging. However, studies mainly focus on single scanners or specific body areas, indicating a need for broader evaluations. Our study uses a Dose Monitoring System (DMS) to compare vertical shifts in CT exams from two scanners, one AI-based and one manually operated. We analysed 6983 CT scans from 3000 patients on two scanners operated by the same radiology team using the GE DoseWatch (GE Healthcare, Milwaukee, USA) platform. Statistical analysis included tests for normality and distribution comparison (p<0.05). Parameter estimation used an iterative bootstrap method. We also evaluated how many scans have vertical displacement greater than 20 mm. Our results showed non-Gaussian vertical shift distributions for both scanners (p <0.01) and significant differences between them (p < 0.01). Notably, 23% of Ascend exams had vertical shifts beyond ±20 mm, compared to 43% for Lightspeed, indicating substantial improvement with AI-assisted positioning. These findings demonstrate that DMSs can measure alignment accuracy, allowing for the comparison of positioning protocols. However, 23% of AI-assisted examinations showed misalignment, highlighting the need for ongoing training and oversight for technical staff, especially in complex cases. Limitations include reliance on specific software and the absence of image quality and radiation dose comparisons. Future studies should analyse the longterm performance of AI in various clinical settings. This study highlights the importance of continuous data analysis for monitoring system performance and identifying training needs. Better positioning accuracy enhances patient care. The study suggests that while AI-based positioning systems provide substantial benefits, their practical use depends on careful integration into clinical workflows and ongoing training of technical staff.