Related Experiment Video
Updated: May 22, 2025
Novel In Vivo Micro-Computed Tomography Imaging Techniques for Assessing the Progression of Non-Alcoholic Fatty Liver Disease
Published on: March 24, 2023
Diagnosis of moderate-to-severe hepatic steatosis using deep learning-based automated attenuation measurements on
Hae Young Kim1, Kyung Jin Lee1, Seung Soo Lee2
1Asan Medical Center, University of Ulsan College of Medicine, Seoul, Republic of Korea.
Purpose:
To evaluate the utility of deep learning-based automated attenuation measurements on contrast-enhanced CT (CECT) for diagnosing moderate-to-severe hepatic steatosis (HS), using histology as reference standard.
Methods:
This retrospective study included 3,620 liver donors (2,393 men and 1,227 women; mean age, 31.7 ± 9.4 years), divided into the development (n = 2,714) and test (n = 906) cohorts. Attenuation values of the liver and spleen on CECT were measured both manually and using a deep learning algorithm (before and after radiologists' correction of segmentation errors). Performance of: (1) liver attenuation and (2) liver-spleen attenuation difference for diagnosing moderate-to-severe HS (> 33%) was assessed using the area under the receiver operating characteristic curve (AUC). Three different criteria targeting 95% sensitivity, 95% specificity, and the maximum Youden's index, respectively, for diagnosing moderate-to-severe HS, were developed and validated.
Results:
The performance of deep learning-based measurements did not differ significantly, with or without radiologists' corrections (p = 0.13). Liver-spleen attenuation difference outperformed liver attenuation alone in diagnosing moderate-to-severe HS in both deep learning-based (AUC, 0.868 vs. 0.821; p = 0.001) and manual (AUC, 0.871 vs. 0.823; p = 0.001) measurements. In the test cohort, the criterion targeting 95% sensitivity for diagnosing moderate-to-severe HS (liver-spleen attenuation difference ≤ 2.8 HU) yielded 92.0% (69/75) sensitivity and 48.5% (403/831) specificity. The criterion targeting 95% specificity (liver-spleen attenuation difference ≤ -18.8 HU) yielded 53.3% (40/75) sensitivity and 95.7% (795/831) specificity. The criterion targeting the maximum Youden's index (liver-spleen attenuation difference ≤ -8.2 HU) yielded 82.7% (62/75) sensitivity and 80.7% (671/831) specificity.
Conclusion:
Deep learning-based automated measurements of liver and spleen attenuation on CECT can be used reliably to detect moderate-to-severe HS.
