Comparison of Deep Learning Architectures in Classification of Microcalcifications Clusters in Digital Mammograms

Microcalcifications clusters (MCCs) are relevant breast cancer indirect evidence and early detection can prevent death. In this paper, we carry out a comparison of the deep learning architectures (DL) InceptionV3, DenseNet121, ResNet50, VGG-16, MobileNet V2, LeNet-5 and AlexNet in classification of MCCs in digital mammograms, with the aim to select the best configuration and building blocks that yield a reduced number of parameters. We used the INbreast database to extract patches of size 144x144 pixels corresponding to 1 cm(2). The networks were implemented and trained using four independent configurations that consisted of: training each architecture without any extra layer, by adding, after each convolutional and fully connected layer, a Batch Normalization layer, an L2 regularization layer or a Dropout layer, respectively. The best overall accuracy was yielded by the MobileNetV2 with Dropout configuration. The network is built of residual blocks, depth separable convolutional blocks and 1x1 convolutional layers this allows a reduced number of parameters while yielding an accuracy of 0.99841. The comparison highlights the differences and similarities of the DL networks necessary to make the best decision possible to design and implement more optimal networks to classify MCCs. Furthermore, it demonstrates that for the classification of these lesions, the shallow architectures built with certain building blocks and trained with the right configuration, give similar results to their deeper and more complex counterparts.