|ECR 2019 / C-3419|
|Machine Learning-Assisted Prediction of Hepatic Steatosis using 3D-CNN Auto-segmentation of Contrast-enhanced Portal Venous Phase CT Examinations of the Abdomen|
In our study, a hybrid 3D-CNN/statistical approach is able to predict Average Plain Density of a liver from contrasted examinations at a level comparable with that of human radiologist’s assessment to within 4-5HU from standardized measurements of Uncontrasted CT. The use of Plain CT as an internal reference standard provided sufficient data for training both Unenhanced and PV DCNN’s to identify hepatic steatosis to the degree possible from Unenhanced CT. Of note, this task is performed in a fully-automated fashion with zero user interaction from the interpreting radiologist requiring neither segmentation or user input.
The successful development of this fully automated, trained algorithm may allow for the quick, objective, and validated confirmation of steatosis on routine CT. Fundamentally, this network could provide value-added information to a reporting radiologist for any contrast CT abdomen case interpreted, who would then be free to agree or disagree with the conclusions drawn from the algorithm using routine clinical radiology analysis. The results provided by the AI would help direct the attention of an interpreting radiologist, particularly in milder or less obvious cases of hepatic steatosis as well as improve the sensitivity and specificity of Portal Venous characterization of steatosis, something that can be conventionally quite difficult or tedious to perform in everyday clinical radiological workflow.
Thematically related posters
ECR 2019 / C-0289
Radiomic analysis in hepatocellular carcinoma to predict clinical outcome after liver transplantation: a preliminary study
ECR 2019 / C-3569
Radiographic diagnosis of cervical ossification of the posterior longitudinal ligament using deep neural network