Breast cancer: deep transfer learning techniques for breast tumor detection in mammography
Saida Sarra Boudouh, Mustapha Bouakkaz · IET conference proceedings. · 2023
Among women, breast cancer become the second leading cause of death. Several techniques exist for breast cancer detection, yet mammogram imaging is currently the most often-used method. To avoid overfitting challenges with images, we were able to meet the goal of our study, which was to construct an accurate Convolutional Neural Network (CNN) model that classifies mammography images into normal and abnormal using deep transfer learning and data augmentation techniques. Xception, ResNet50V2, and VGG16 were used with, then without trainable layers on the Mammographic Image Analyses Society MiniMammographic Database (MiniMIAS), in order to determine which one of them is suitable for our case. But due to the shortage of images in the abnormal class and to avoid unbalanced data, we managed to balance the data using images from the Chinese Mammography Database (CMMD) which only contains abnormal images. The dataset was pre-processed using various filters to extract the Region Of Interest (ROI) and remove any noises, resulting in better images for the training process. The evaluation results show that using this dataset the pre-trained models Xception and VGG16 with trainable layers are more suitable than ResNet50V2 for our situation with an accuracy of 99.9% ,95.23%, and 94.04% respectively, which proves that the pre-processing filters employed are extremely effective in breast cancer detection. With this dataset, we were also able to demonstrate that activating the layers of a pre-trained model to train will produce more positive results than not doing so for breast tumor detection.