Exploring the Best Parameters of Deep Learning for Breast Cancer Classification System

Abstract

Breast cancer is one of the deadliest cancers in the world. It is essential to detect the signs of cancer as early as possible, to make the survival rate higher. However, detecting the signs of breast cancer using the machine or deep learning algorithms from the diagnostic imaging results is not trivial. Slight changes in the illumination of the scanned area can significantly affect the automatic breast cancer classification process. Hence, the research aims to propose an automatic classifier for breast cancer from digital medical imaging (e.g., Positron Emission Tomography or PET, X-Ray of Mammogram, and Magnetic Resonance Imaging (MRI) images). The research proposes modified deep learning architecture with five different settings to model automatic breast cancer classifiers. In addition, five machine learning algorithms are also explored to model the classifiers. The dataset used in the research is the Curated Breast Imaging Subset of Digital Database for Screening Mammography (CBIS-DDSM). A total of 2,676 mammogram images are used in the research and are split into 80%:20% (2,141:535) for training and testing datasets. The results demonstrate that the model trained with eight layers of Convolutional Neural Networks (CNN) (SET-8) achieves the best accuracy score of 94.89% and 93.75% in the training and validation dataset, respectively.