2.2. Image Preprocessing

Sample images of the DDSM dataset of each category are shown in Figure 1, consisting of squared ROI images of different sizes. These images were converted to a standard size of 299 × 299, using inter-cubic interpolation algorithms, as the convolutional neural network (CNN) structure operated on images of the same size.

(a) Original mammogram images taken from DDSM dataset; (b) cropped ROI images of original images; (c) resized ROI images of size 299 × 299.

DDSM mammography data were single-channel image sets, different from natural color images applied in the pretrained models. However, basic image features in terms of edges, shapes, and other high-level features can still be extracted using pretrained convolutional neural network-based models. In this respect, we note that the sample DDSM image was a single-channel image, whereas the input to the InceptionResNetV2 model required a three-channel image. Thus, sample DDSM images were converted to three channels by copying the pixel value of single-channel images to the other two channels.