Accuracy of Aneurysm Model Reconstruction: Influence of Software, Thresholding, and Operator Variability

Accurate reconstruction and evaluation of intracranial aneurysm (IA) models are essential for clinical diagnosis, treatment planning, and biomedical researches such as morphology biomarkers assessment and hemodynamic analysis using computational fluid dynamics (CFD) simulations. The main objective of this study is to quantify the impact of segmentation thresholds and software platforms on the reconstruction accuracy of IA geometry as well as the impact of inter-user variability on the assessment of IA geometry. 600 aneurysm models were reconstructed from 100 patient-specific digital subtraction angiography (DSA) images using Materialise Mimics and 3DSlicer at three grey value (GV) thresholds; 1000, 1500, and 2500. Geometric measurements were performed in 3-matic by three researchers (R1, R2, and R3) with different experience levels. Several vessel diameters and aneurysm morphology parameters were measured independently. Data from Mimics with the 2500 GV threshold, and the most experienced user (R1) served as baselines for comparison. Normality was evaluated using Shapiro-Wilk tests, and statistical differences were assessed with paired t-tests and relative percent differences. All anatomical regions showed statistically significant geometric variation across software and threshold. Model evaluation showed potential statistically significant variation between users. Models from 3DSlicer were consistently smaller than those from Mimics with percentage differences ranging from −1.27% to −4.38% (all p < .05). Lower thresholds produced consistently larger models; decreasing from 2500 GV to 1000 GV increased average diameters by up to 15.9%, depending on specific region (p < .05 for all comparisons). User-related variability was most pronounced in the least experienced user (R3), with size measurements deviating by up to 22.67% from R1 (p = 2.36×10⁻¹⁷), while R2's measurements showed minimal differences. The geometry and evaluation of IA models vary based on the software, threshold choices, and user. The results show that standardizing protocols for threshold values and operator training, alongside implementing transparent and validated workflows, is critical for both clinical and research applications.