Benchmarking of ResNet models for breast cancer diagnosis using mammographic images

Abstract

Breast cancer is one of the cancer types with a high mortality rate worldwide. Early diagnosis is of great importance to reduce this mortality rate. Computer-aided early diagnosis systems enable doctors to make more precise and faster decisions. The Mammographic Image Analysis Society (MIAS) dataset was used in this study. The breast area was selected by masking in mammography images. The number of images was increased using data augmentation techniques. Mammography images were classified as normal, benign and malignant using four different ResNet models. The highest classification accuracy was achieved by using ResNet18 model with 93.83%. The accuracies obtained with ResNet50, ResNet101 and ResNet152 were 87.24%, 87.44% and 91.25% respectively.