Figure 3

CNN Training and Validation. Performance of the CNN on the discovery dataset: (a) accuracy, (b) cross entropy and (c) improvement (Δ cross entropy). Improvement shown in panel (c), in computed on the test set only, preventing the model from overfitting. Performance of the CNN on the validation dataset: (d) the Area under the ROC curve (AUC) of the probability map with respect to the reader segmentation, and (e) Dice Similarity Coefficient (DSC) of the generated segmentations.