Classification of Breast Cancer Images by Transfer Learning Approach Using Different Patching Sizes

Abstract

Cancer is the second leading cause of death worldwide, and breast cancer is the fifth leading cause of death. Therefore, early diagnosis is a vital stage. The diagnosis of this cancer is usually made with histopathological images. In this study, four classes of breast cancer images (normal, in-situ/regional, benign, and invasive) were detected by using a transfer learning approach. The BACH data set was used in this study. The images in the data set were divided into small patches of 256 x 256, 128 x 128, and 75 x 75 sizes from the places where the cell density is high. In addition, the bottleneck feature was used, which provides higher performance in low image numbers. In this way, the bottleneck feature performances of the models were investigated, and it was aimed to get a high accuracy rate with the processes performed in the pre-processing steps. As a result, the success of our system is 99% accuracy in 256 x 256 patch size and it varies according to the size of the image patches.