Daiju Ueda and colleagues at Osaka Metropolitan University developed a deep learning-based model that effectively detects fatty liver disease (steatosis) in chest radiographs. The study used data from 4,414 patients from two Japanese hospitals, who underwent 6,599 chest radiographs and liver elastograms with controlled attenuation parameter (CAP, a quantitative measure of steatosis). Patients from one of the clinics were randomly assigned in an 8:1:1 ratio to the training, tuning, and internal testing datasets of the model, while participants from the second clinic were included in the external testing dataset. The results are published in the journal Radiology: Cardiothoracic Imaging.
The dataset for internal testing included 529 chest radiographs of 363 patients (mean age 56 years; 344 men), for external testing — 1100 radiographs of 783 patients (mean age 58 years; 604 men). During internal testing, the area under the ROC curve was 0.83; accuracy — 77 percent; sensitivity — 68 percent; and specificity — 82 percent. During external testing, these indicators were 0.82; 76, 76, and 76 percent, respectively. The model performance was rated as good.
