Figure S1: Methodology for data partitioning. Left branch: Data were split into Training + Validation (85%) and Test (15%). Then, the Training + Validation was used for five-fold cross-validation (85% for training, 15% for validation), and all sets have similar proportion of steatosis grades. Results from the cross-validation settings were used to define the best model configuration, in terms of architecture and hyperparameters. Right branch: The dataset used for readers’ evaluation contained 52 patients, we used the same patients as a new test set, named “readers set.” The two test sets were evaluated using the best configuration, but each was trained on its respective training data.

Figure S2: Receiver operating characteristic analysis of each reader for classification of dichotomized histology-determined steatosis grades, S0 versus ≥ S1, ≤ S1 versus ≥ S2, and ≤ S2 versus S3. Each letter from A to F denotes a different reader, 1 and 2 denotes the reading session.

Figure S3: Confusion matrix for the dichotomized steatosis grades, S0 versus ≥ S1, ≤ S1 versus ≥ S2, and ≤ S2 versus S3, using the deep learning model and the readers set.