An Ad Hoc Random Initialization Deep Neural Network Architecture for Discriminating Malignant Breast Cancer Lesions in Mammographic Images

Figure 3

Example of a batch of 16 images from the training set. The ROI from which each image has been generated has been randomly rescaled (independently over the two axes), rotated by a random angle, randomly flipped, and resampled to fit into a pixel frame with aspect ratio 1. Any remaining area not filled by the image is padded with an array of pixels drawn from the edge of the image.