A 1631 mediolateral oblique (MLO) mammogram dataset was obtained from an open database [6]. Mammography positioning is evaluated using guideline criteria in the pectoral major muscle, retromammary space, bilateral breast symmetry, IMF, and nipple area. The retromammary space was targeted in our study.
The labeled mammograph's region of interest (ROI) was automatically extracted from the retromammary space. After the grayscale mammogram images were converted to binary images, morphological filters were applied to the images to eliminate radiopaque artifacts and labels. To remove the shading, binary masking was applied using a rectangular mask. Morphological operations were performed on the binary images. The left-bottom extrema points and x- and y-coordinates of one of the points were detected using the measured properties of the breast regions. An ROI with a size of both 750 and 1500 pixels height was set around the extrema point as an index.
The extracted ROIs were classified into three classes (excellent, average, and poor) using five representative DCNN models. The VGG-16, Inception-v3, ResNet50, Inception-ResNet-v2, and Xception were pre-trained using ImageNet to perform transfer learning. These methods were implemented using Python 3.6, TensorFlow 1.15, and Keras 2.1. They were then evaluated in an environment with Windows 10 OS and an NVIDIA GeForce GTX 1080 Ti GPU. To train the models, the maximum number of epochs was set to 50. The batch size was set to 16 using the adaptive moment estimation (Adam) optimizer. The classification accuracy was determined from the resulting confusion matrix. Several metrics were used to evaluate the performance of the DCNN: accuracy, class activation map, and softmax value.
