ERS3026. Diffuse Hepatic Steatosis on Unenhanced CT: Accuracy of Existing Diagnostic Criteria and the Modeling of Novel Ones
Authors* Denotes Presenting Author
Ernest Wang *;
Pennsylvania State University College of Medicine
Benjamin Shin;
George Washington University Hospital
Kathryn McGillen;
Pennsylvania State University College of Medicine
Objective:
The diagnosis of diffuse hepatic steatosis on unenhanced CT frequently is defined as meeting either absolute liver hypoattenuation less than 40 HU, or relative hypoattenuation with a spleen-minus-liver difference of at least 10 HU (Hamer, et al. 2006). These criteria were based on early studies with smaller cohorts to identify moderate-to-severe steatosis (Piekarski, et al. 1980; Yajima, et al. 1982; Limanond, et al. 2004; Kodama, et al. 2007). In this ongoing retrospective study our aims were to: A) re-test the associations of liver and spleen attenuations with pathology steatosis grading in a larger cohort; B) re-examine the accuracy of existing criteria given recent advances in CT; and C) determine whether decision tree classification modeling would yield novel and improved criteria.
Materials and Methods:
This study drew from a sequential list of patients who underwent liver biopsy at a tertiary center. Records were reviewed in reverse chronologic order to obtain 300 subjects with unenhanced CT of the abdomen performed within 12 months of liver biopsy. Liver and spleen regions-of-interest were placed on axial sections by a blinded investigator to obtain mean attenuations. The relationships of liver and spleen attenuations, their difference, and pathology steatosis grading (none/insignificant: < 5%; mild: 5-33%; moderate: > 33-66%; severe: > 66%) were tested using Spearman correlations. Then, sensitivities and specificities of the existing criteria were calculated for the present cohort. Finally, attenuation features were employed in decision tree classification models with cross-validation and pruning to arrive at novel criteria balancing parsimony with sensitivity and specificity.
Results:
Pathology steatosis grading associated with both liver attenuation (rho = -0.55; p < 0.001) and the spleen-liver difference (rho = 0.55; p < 0.001), but not with spleen attenuation (rho = -0.03; p = 0.614). The sensitivity and specificity of absolute liver hypoattenuation (< 40 HU) to detect moderate-to-severe steatosis were 73% and 75%, respectively, and 39% and 96% using relative liver hypoattenuation (spleen-liver difference at least 10 HU). In classification modeling to detect moderate-to-severe steatosis, the single-node model with the lowest misclassification rate (17%) used a liver attenuation threshold of < 36.3 HU, with a sensitivity and specificity of 60% and 89%. For the detection of any degree of steatosis, the two-node model with the lowest misclassification rate (25%) used: Node 1) a liver attenuation threshold of < 40.2 HU, and Node 2) for those not meeting the first criteria, a spleen-liver difference > 8.7 HU, for a sensitivity and specificity of 88% and 59%.
Conclusion:
Both absolute and relative liver hypoattenuation demonstrate strong associations with pathology-based steatosis grading. Incorporating both criteria in a decision tree model instead of using them as “either-or” criteria may harness their relative strengths to identify any degree of steatosis. This preliminary data is encouraging and we expect future work with a larger cohort to validate more precise thresholds.
