Abstract
ABSTRACT
Purpose
To quantify the metrological sampling error of conventional 2D regions of interest (ROIs) relative to whole-organ 3D volumetric CT attenuation ground truth in opportunistic screening for hepatic steatosis, and to evaluate its functional dependency on intraparenchymal spatial heterogeneity.
Materials and Methods
In this retrospective computational investigation adhering to CLAIM and TRIPOD-AI reporting standards, 971 non-contrast and contrast-enhanced abdominal CT examinations were analyzed across a multicenter derivation cohort (TotalSegmentator v2, n = 841) and an external validation cohort (Medical Segmentation Decathlon Task03 Liver, n = 130). Automated 3D liver segmentation was conducted using nnU-Net, supplemented by an 8-mm spherical 3D morphological erosion to eliminate peripheral partial volume effects, liver capsule artifacts, and major hilar vascular structures. Three clinical sampling paradigms (1, 3, and 5 axial circular 2D ROIs of 15 mm diameter) were simulated through 500 stochastic Monte Carlo iterations per examination (485,500 total simulations). Absolute metrological sampling error (|ΔHU| = |HU_ROI - HU_3D|) and diagnostic steatosis misclassification rates at the standard unenhanced threshold (HU_3D ≤ 40 HU) were evaluated. Statistical significance across sampling paradigms was tested using two-tailed paired Wilcoxon signed-rank tests.
Results
In the derivation cohort (n = 841), the conventional guideline-recommended 3-ROI paradigm yielded a mean absolute sampling error of 6.61 ± 7.83 HU (median: 2.92 HU; interquartile range: 1.34–8.45 HU). While representing a 41.6% error reduction compared to the single-ROI approach (11.32 ± 14.16 HU; W = 0, p
Conclusion
Manual 2D ROI sampling exhibits an inherent physical uncertainty that exceeds the 5-HU clinical tolerance threshold in parenchymally heterogeneous livers, misclassifying nearly half of borderline examinations. Accurate opportunistic steatosis screening requires transition to automated whole-organ 3D volumetric CT densitometry embedded in PACS networks.
KEY POINTS
Conventional 3-ROI sampling leaves an average metrological error of 6.61 ± 7.83 HU compared with 3D whole-liver ground truth.
Intraparenchymal spatial heterogeneity is the primary determinant of 2D sampling error (r = 0.850, p
In borderline diagnostic attenuation ranges (35–45 HU), 2D manual sampling misclassifies up to 45.2% of patients with spatial heterogeneity.