Figure 1.

Workflow of data analysis. The entire workflow consists of three steps. In step 1, we tune the number of layers to freeze in order to do TL optimally. Our experiments show that when freezing 7 layers, our model achieved the best predictive performance. We then test the model using a fivefold grouped cross-validation. Finally, for each input photo and the corresponding model prediction, we generate a saliency map for each anomaly (regardless of presence in the photo).