Breast Cancer: Breast Tumor Detection Using Deep Transfer Learning Techniques in Mammogram Images

Breast cancer has become the second leading cause of death among women. Mammogram images are the best and most often used method for detecting breast cancer. This study aims to build an accurate breast tumor detection model that classifies mammogram images into normal and abnormal (tumor). Transfer learning and data augmentation approaches were employed to prevent imagery overfitting problems using three pre-trained convolutional neural networks: AlexNet, VGG16, and VGG19 without trainable layers. The dataset was gathered from two different databases to ensure diversity in the images. We gathered 4000 images from each Digital Database for Screening Mammography and the Chinese Mammography Database. The obtained results were very satisfying according to the previous studies. 99.5% and 100% accuracy were obtained from AlexNet and VGG16 (with trainable layers), better than the existing breast tumour detection models. The applied filters in the pre-processing phase were effective based on the obtained results. Also, we determined that adopting a pre-trained model with trainable layers provides better results than using the model as it is without training the layers. Breast cancer has become the second leading cause of death among women. Mammogram images are the best and most often used method for detecting breast cancer. This study aims to build an accurate breast tumor detection model that classifies mammogram images into normal and abnormal (tumor). Transfer learning and data augmentation approaches were employed to prevent imagery overfitting problems using three pre-trained convolutional neural networks: AlexNet, VGG16, and VGG19 without trainable layers. The dataset was gathered from two different databases to ensure diversity in the images. We gathered 4000 images from each Digital Database for Screening Mammography and the Chinese Mammography Database. The obtained results were very satisfying according to the previous studies. 99.5% and 100% accuracy were obtained from AlexNet and VGG16 (with trainable layers), better than the existing breast tumour detection models. The applied filters in the pre-processing phase were effective based on the obtained results. Also, we determined that adopting a pre-trained model with trainable layers provides better results than using the model as it is without training the layers.