Deep Learning for Segmentation-based Hepatic Steatosis Detection on Open Data: A Multicenter International Validation Study
share
Despite high global prevalence of hepatic steatosis, no automated diagnostics demonstrated generalizability in detecting steatosis on multiple heterogeneous populations. In this retrospective study, we externally validated a fully automated artificial intelligence (AI) system to detect hepatic steatosis. 1,014 non-contrast enhanced chest computed tomography (CT) scans were collected from eight distinct datasets: LIDC-IDRI, NSCLC-Lung1, RIDER, VESSEL12, RICORD-1A, RICORD-1B, COVID-19-Italy, and COVID-19-China. This three-step AI workflow consists of the following: (i) 3D liver segmentation - a 3D U-Net deep learning model developed for liver segmentation and applied externally without retraining. (ii) liver attenuation measurements by three automatic methods: AI on regions of interest (AI-ROI), AI-3D, and AI-2D; (iii) hepatic steatosis detection. The deep-learning segmentation achieved a mean dice coefficient of 0.957. AI-ROI attenuation measurements showed no significant differences compared to expert measurements (P > 0.05), but AI-3D and AI-2D were significantly different from the expert (P < 0.001). The area under the curve (AUC) of steatosis classification for AI-ROI, AI-3D, and AI-2D are 0.921 (95 CI: 0.883 - 0.959), 0.939 (95 0.938) respectively. If adopted for universal detection, this deep learning system could potentially allow early non-invasive, non-pharmacological preventative interventions for hepatic steatosis. 1,014 expert-annotated liver segmentations of CT images can be downloaded here: https://drive.google.com/drive/folders/1-g_zJeAaZXYXGqL1OeF6pUjr6KB0igJX.
READ FULL TEXT