Summary of the classification performance with confidence intervals (CIs) computed at 95% using bootstrapping (n=1000). “AUC” refer to the receiver-operating curve. “Loc Bag” and “Loc GBP” respectively refer to the localization precision of the sparse BagNet and Guided Backpropagation on ResNet-50 at localizing lesions from annotated images. For each dataset, the first row shows the performance of the interpretable sparse BagNet model, while the second row shows the performance of the baseline black-box ResNet-50 model. The Kaggle dataset (first row) is the internal dataset used to train and evaluate the model, while the other datasets were used for external validation to assess the generalization properties of the trained model. The low classification performance on the FCM-UNA and FGA-DR datasets can be explained by the relatively low quality of most images in the FCM-UNA dataset and the large intensity variation of the FGA-DR dataset (S

Summary of the classification performance with confidence intervals (CIs) computed at 95% using bootstrapping (n=1000). “AUC” refer to the receiver-operating curve. “Loc Bag” and “Loc GBP” respectively refer to the localization precision of the sparse BagNet and Guided Backpropagation on ResNet-50 at localizing lesions from annotated images. For each dataset, the first row shows the performance of the interpretable sparse BagNet model, while the second row shows the performance of the baseline black-box ResNet-50 model. The Kaggle dataset (first row) is the internal dataset used to train and evaluate the model, while the other datasets were used for external validation to assess the generalization properties of the trained model. The low classification performance on the FCM-UNA and FGA-DR datasets can be explained by the relatively low quality of most images in the FCM-UNA dataset and the large intensity variation of the FGA-DR dataset (S