Breast Cancer Classification Using Deep Learning

Abstract

Breast cancer continues to pose a significant global health challenge, evident from the staggering statistics of 10 million deaths and 19.3 million new cases worldwide in 2018. Conventional testing methods are costly, prone to human error, and often yield inaccurate results. The classification of breast cancer based on mammography images presents its own set of challenges, exacerbated by shortages of trained radiologists and screening equipment. Ethiopia is particularly alarming due to its highest age-standardized death rate from breast cancer. With a projected 50% increase in cancer-related mortality by 2030, the importance of early detection and precise classification cannot be overstated. In addressing breast cancer detection in Ethiopia, the proposal suggests the implementation of a Convolutional Neural Networks (CNN) model, necessitating collaboration among academics, policymakers, healthcare practitioners, and IT experts. Previous research conducted in Ethiopia has demonstrated CNNs achieving accuracy rates of up to 88% using solely local datasets. In this study, the accuracy of breast cancer classification was assessed using public, local, and combined mammography image datasets, employing two transfer learning architectures: ResNet50 and EfficientNetB0. The datasets were partitioned into three subsets: 70% for training, 15% for validation, and 15% for testing, totaling 2376 mammography images classified as benign or malignant. Applying EfficientNetB0 to the local, public, and merged image datasets yielded an accuracy of 78.43% with ROC curves of 0.81, 0.89, and 0.46 for 40 epochs with a learning rate of 0.0001respectively. Similarly, utilizing ResNet50 on the same datasets resulted in an accuracy of 78.43% with ROC curves of 0.81, 0.69, and 0.81 for 40 epochs with a learning rate of 0.0001respectively. The study conducted a comparison between the transfer learning architectures of ResNet50 and EfficientNetB0 using mammography images obtained from local, merged, and publicly available datasets. Results from the analysis of the local dataset revealed an accuracy of 78.43% with ROC curves of 0.81, indicating that EfficientNetB0 outperformed ResNet50 in accurately classifying breast cancer as either benign or malignant.