Deep-learning segmentation to select liver parenchyma for categorizing hepatic steatosis on multinational chest CT

Abstract

Unenhanced CT scans exhibit high specificity in detecting moderate-to-severe hepatic steatosis. Even though many CTs are scanned from health screening and various diagnostic contexts, their potential for hepatic steatosis detection has largely remained unexplored. The accuracy of previous methodologies has been limited by the inclusion of non-parenchymal liver regions. To overcome this limitation, we present a novel deep-learning (DL) based method tailored for the automatic selection of parenchymal portions in CT images. This innovative method automatically delineates circular regions for effectively detecting hepatic steatosis. We use 1,014 multinational CT images to develop a DL model for segmenting liver and selecting the parenchymal regions. The results demonstrate outstanding performance in both tasks. By excluding non-parenchymal portions, our DL-based method surpasses previous limitations, achieving radiologist-level accuracy in liver attenuation measurements and hepatic steatosis detection. To ensure the reproducibility, we have openly shared 1014 annotated CT images and the DL system codes. Our novel research contributes to the refinement the automated detection methodologies of hepatic steatosis on CT images, enhancing the accuracy and efficiency of healthcare screening processes.

Keywords: Computed tomography, nnU-Net, Deep learning, Hepatic steatosis.

Figure 1

Study design. (a) A fully automated deep learning-based system was designed for measuring live attenuation and assessing moderate-to-severe steatosis. The system consists of three steps: 3D liver segmentation, attenuation measurement, and steatosis classification. The development data consisted of 479 CT images from the LIDC-IDRI dataset, which were split into 4:1:1 for training, tuning, and internal testing, respectively. External validations were conducted on a total of 535 CT images from six datasets across eight countries. (b) The performance of the deep learning system was evaluated using radiologist-validated ground truths. (c) Three deep learning-based methods were compared to measure CT attenuation in Hounsfield units (HU). They are (1) DL-volumetric, which measured attenuation volumetrically; (2) DL-axial, which measured attenuation on a single axial slice containing the largest cross-sectional area; and (3) DL-parenchymal, which measured attenuation on parenchymal portions selected on axial slices using 3D auto-segmentation. Figure c displays examples of DL-based attenuation measurements. DL deep learning, HU Hounsfield unit.

Figure 2

Deep learning segmentation performance. DSC indicates the similarity between expert and DL segmentations. The DSC is also ground by different conditions of the liver: “Normal vs. Steatosis”. The number of CT images is denoted at the bottom (normal vs. steatosis). The result shows that DL achieves an overall high segmentation accuracy, and slightly better performance on the normal liver than that of steatosis. DL = deep learning, DSC = dice similarity coefficient. ***, p < 0.001; **, p < 0.01; *, p < 0.05.

Figure 3

Concordances of DL-parenchymal vs. ground-truth attenuation. DL-parenchymal attenuation was compared to ground-truth attenuations on partitioned datasets (development set: n = 399, internal test set: n = 80, external test set: n = 535). The scatter plot shows the p-value, slope, and Spearman correlation coefficient. The agreement was further assessed using Bland–Altman analysis, with liver attenuations in Hounsfield units (HU) on the x-axis. The bold line represents the actual mean difference (error), and the other two dotted lines show 95% limits of agreements.

Figure 4

Examples of DL-parenchymal vs. ground-truth attenuation. Two examples presented the portion placement of DL-parenchymal attenuation. In Figure a. 1, manual segmentation and selection were used to measure the liver attenuation of a no steatosis condition. In contrast, DL auto-segmentation and DL-selected parenchymal regions were used for attenuation measurements in Figure a. 2. Another example of the liver with steatosis was shown in Figures b. 1 and b. 2. DL deep learning, DSC Dice similarity coefficient.

Figure 5

Attenuation concordance of multiple human experts and DL-based methods. A subset of 100 images was independently measured by four human experts and DL-based methods. Significant differences are shown with denoted p values: ***, p < 0.001; **, p < 0.010; *, p < 0.050. DL deep learning.