EP130 AI-oriented nerve segmentation: more methodology for more consistency?

<h3>Background and Aims</h3> Artificial intelligence (AI) promises to enhance performance of Ultrasound-guided regional anesthesia (UGRA). However, existing methodologies seems to lack consistency, which could compromise reliability and clinical applicability. <h3>Methods</h3> A prospective observational study was conducted involving 54 adult patients scheduled for elective carpal tunnel surgery. 1) Validity of ultrasound measurement was assessed by comparing surgical and ultrasound measurement of median nerve diameters. 2) Reliability was assessed by comparing the cross-section area of the median measured by ultrasound independently by 3 experts applying rigorous radiological segmentation criteria. Inter-observer agreement was assessed using different metrics (See figure 1) <h3>Results</h3> 1) No significant difference was observed between ultrasound (mean diameter: 6.59 mm – SD: 1.16) and surgical measurements (mean diameter: 6.62 mm, SD = 1.02, p = 0.833). 2) Inter-observer variability was minimal, demonstrated by high ICC values (0.95), and excellent segmentation overlap metrics: median(IQR) DSC = 0.94 (0.02), IoU = 0.89 (0.03), precision = 0.94 (0.03), recall = 0.95 (0.03), and low Hausdorff distance (0.29 mm, IQR: 0.11). <h3>Conclusions</h3> The consistently high inter-observer reliability suggests that standardised, rigorous segmentation criteria defining the nerve contour based on radiological practice will significantly reduce inter-observer variability. These findings should be confirmed on larger sequences at other anatomical sites. Standardised segmentation procedures do not exist and should be developed. AI-based nerve segmentation for UGRA can achieve good validity and reliability through standardized methodologies. Ensuring high-quality data acquisition and rigorous segmentation criteria optimizes clinical relevance, ultimately enhancing patient safety and procedural accuracy.