A deep learning architecture for identification of breast cancer on mammography by learning various representations of cancerous mass

Abstract

Deep Learning (DL) is a high capable machine learning algorithm with the detailed analysis abilities on images. Although DL models achieve very high classification performances, the applications are trending on using and fine-tuning pre-trained DL models by transfer learning due to the dependence on the number of data, long train time, employments in modeling the most meaningful architecture. In this chapter, we proposed own pruned and simple DL architectures on ROIs extracted from mammography to classify cancer-normal using Convolutional Neural Network (CNN) and Deep Autoencoder (Deep AE) models, which are the most popular DL algorithms. Breast Cancer, which occurs as a result of developing of normal breast tissue to a tumour, is one of the deadliest diseases according to WHO reports. The detection of cancerous mass at early stages is one of the decisive step to start the treatment process. Mammography images are the most effective and simplest way of the diagnosis of breast cancer. Whereas early diagnosis of breast cancer is a hard process due to characteristics of mammography, the computer-assisted diagnosis systems have ability to perform a detailed analysis for a complete assessment. The aim of this study is proposing a robust cancer diagnosis model with a light-weighted DL architecture and comparing the efficiency of the dense layer with the Deep AE kernels against CNN. The ROIs from mammography images were fed into the DL algorithms and the achievements were evaluated. The proposed Deep AE architecture reached the classification performance rates of 95.17%, 96.81%, 93.65%, 93.38%, 96.95%, and 0.972 for overall accuracy, sensitivity, specificity, precision, NPV, and AUROC, respectively.