Figure A5.

Comparisons of accuracies of the pre-trained models and the proposed model appendix by testing on validation dataset 2. A transfer learning method was used to apply other state-of-the-art pre-trained models to predict the malignancy and invasiveness of GGNs. In the pre-trained model development, the parameters in CNN pooling processes were frozen and the parameters in fully connected or classifier layers were trained. In the model training process, the pre-trained models were the configured with the same parameters as the DNN model.