Breast cancer diagnosis with deep learning methods from mammogram images

Early detection and diagnosis is very important to reduce the mortality rate in breast cancer. For this reason, it is necessary to accurately analyze mammogram images, one of the breast cancer imaging methods. Various methods have been developed to achieve high accuracy and faster classification of mammogram images. The aim of this study is to provide high accuracy and speed by using deep learning methods for the classification of mammogram images. In order to achieve high accuracy rate, image filtering and histogram equalization methods can be used in the image processing step before classification using deep learning methods. In this study, 20000 mammogram images from DDSM, Inbreast, MIAS datasets were used. The deep learning methods used are VGG16, VGG19, ResNet50. The image filtering methods used are median filter, binary filter, averaging filter, gaussian filter. In addition, classical histogram equalization and adaptive histogram equalization with contrast limit were used for histogram equalization. Transfer learning and fine-tuning methods were applied to construct the network, resulting in high accuracy. The best result was 99.93% accuracy. This accuracy is very high for the classification of mammogram images. The methods that provide this accuracy value are VGG16 transfer learning network with the use of averaging filter applied to the mammogram image, VGG16 transfer learning network with the use of median filter and classical histogram equalization method applied to the mammogram image, VGG19 transfer learning method with the use of median filter applied to the mammogram image. The results show that histogram equalization and image filtering methods have a positive effect on the classification of mammogram images with deep learning methods.