Real-Time Ultrasound Segmentation of the Inferior Vena Cava: A Study on Software Reliability and Measurement Repeatability

The inferior vena cava ultrasound (IVC) is a non-invasive method used to estimate the patient’s volemic status and support in fluid therapy. A frequently adopted technique to assess fluid responsiveness is the passive leg raise (PLR) maneuver, which improves venous return through lower limbs. However, the reliability of ultrasound measurements depends on the operator’s experience, acquisition conditions, vein motion and non-uniform pulsatility. In this context, the aim of this study was to evaluate the reliability and repeatability of IVC ultrasound measurements performed using VIPER real-time segmentation software, which provides the diameter, caval index (CI), cardiac caval index (CCI) and respiratory caval index (RCI). The analysis focused on evaluating intra and inter-operator variability between measurements obtained with and without real time feedback (RealTime vs NOF), and with the manual method compared to the semi-automatic approach without feedback (Manual VS NOF) for supine and PLR, so as to also compare the ability of the three methods to discriminate between the two conditions. Manual measurements were performed via a graphical user interface (GUI), developed in Python, which allowed comparison with the method traditionally used by clinicians. An experimental protocol was defined which consisted in the acquisition of IVC longitudinal section ultrasound from two nonexpert operators, with and without visual software support, in two different conditions (Supine and PLR) repeating the measurement twice (R1 and R2). The sample consisted of 26 healthy subjects (15 females and 11 males), aged 25 ± 1.37 years. Intra and inter-operator variability was assessed by calculating the Coefficient of Variation (CoV) and Intraclass Correlation Coefficient (ICC) on all measured parameters: mean diameter, CI, CCI, and RCI. The results showed for the RealTime case that all parameters present a lower dispersion than those obtained by post-processing of acquisitions without feedback. CI, RCI and CCI had slightly lower median values of CoV, while diameter's CoVs did not vary significantly. (CI's median CoV, Supine (RealTime): OA=0.15, OB=0.21, inter=0.16; (NOF): OA= 0.15, OB= 0.25, inter= 0.18;). Minor variability was obtained compared to the manual method by observing lower median values of CoV (Manual, Supine CI's median CoV OA=0.24, OB=0.34, inter=28). A linear mixed effects model was developed and showed that the condition had a statistically significant effect on all parameters while the presence of real-time feedback on all but the diameter. To assess to what extent the two conditions could be distinguished using each of the three measurement methods, Cohen’s d was calculated as a measure of effect size. The no-feedback (NOF) approach showed higher discriminative power compared to the manual method. It also outperformed the real-time modality. The%se unexpected results may be explained by a learning bias, as the operators gained experience while simultaneously becoming familiar with the software. These findings highlight the need to repeat the study with clinically experienced operators in order to isolate the true effect of real-time feedback from operator learning effects. Overall, it is evident that the software enables a faster and more standardized acquisition of measurements, with reduced intra- and inter-operator variability, in contrast to the manual method, which remains more time-consuming and operator dependent.