Figure 1. The conceptual architecture of the proposed deep learning CAD framework
Figure 2. The conceptual architecture of the proposed deep learning CAD framework
Figure 4. ROC curves of GoogLeNets trained and tested on the images without a margin(black) and with a margin(red).
Figure 5. Saliency map examples that shows where the important information exists in the image.
Figure 6. ROC curves of GoogLeNet without the data augmentation and with the dataaugmentation.
Figure 7. The perturbation in the center location by radiologist does not affect the performance much
Figure 8. An implementation examples of (a)a possibly benign lesion, (b)a possibly malignant lesion.
Figure 3. (a) and (b) are examples of segmented boundaries of breast lesion.
Table 1. Overview of the lesion size attributes of training data and test data.
Table 2. Comparison of the required time for learning and the number of training images.
Table 3. Performance comparison of a CNN trained on the images with a margin to a network trained on images without a margin. GLN refers to GoogLeNet.
Table 4. Diagnostic performances of CNN networks.
Table 5. Diagnostic performances on centered images, and perturbed images.