Comparative Experimental Investigation and Application of Five Classic Pre-Trained Deep Convolutional Neural Networks via Transfer Learning for Diagnosis of Breast Cancer

In this study, for the diagnosis and classification of breast cancer, we used and applied five classical pre-trained deep convolutional neural network models (DCNN) which have proven successful many times in different fields (ResNet-18, AlexNet, GoogleNet and SuffleNet). To make pre-trained DCNN models suitable for the purpose of our study, we updated some layers according to the new situation by using the transfer learning technique. We did not change the weights of all layers used in these five pre-trained DCNN models. Instead, we just gave new weights to the new layers so that new layers adapt faster to emerging new DCNN models. With these five pre-trained DCNN models, we have realized a quadruple classification as "cancer", "normal", "actionable" and "benign", and a binary classification as "actionable + cancer" and "normal + benign". With these two separate classification and diagnosis studies, we have carried out comparative experimental examination and analysis of pre-trained DCNN models for breast cancer diagnosis. In the study, it was concluded that successful results can be achieved with pre-trained DCNN models without extra time-consuming procedures such as feature extraction, and DCNN can perform quite successfully in cancer diagnosis and image comment.