Deep Convolutional Neural Network Models for Early Detection of Breast Cancer from Digital Mammograms
A. Alice Nithya, P. Shanmugavadivu · Auerbach Publications eBooks · 2024
The most familiar cancer that predominantly victimizes women is breast cancer, and the statistics confirm its rise every year. Hopefully, if breast cancer is found and treated in its early stages, there is a good possibility of recovery. Computer-aided detection/diagnosis (CAD) systems are designed to support the radiologist in diagnosing breast cancer at an early stage. Digital mammography is a commonly trusted medical imaging methodology for breast cancer detection/diagnosis. Recent advancements in deep learning (DL) techniques and their remarkable performance have encouraged several researchers to apply DL techniques for breast cancer detection/diagnosis. The deep convolutional neural networks (DCNN) are proven to effectively handle mammograms to affirm the presence/absence of breast cancer, on par with human intelligence. In certain cases, it is observed that the precision of diagnosis by DCNN outperforms human expertise. The undertaken research study will investigate the potential of various pre-trained convolutional neural network (CNN) architectures, namely VGG19, ResNet50V2, InceptionResNetV2, DenseNet201, and EfficientNetB6, for mammogram classification as either normal or abnormal. The mammogram images of Digital Database for Screening Mammography (DDSM) and INbreast datasets were selected for this study. Further, this chapter comprehends the experimental analysis of proposed DL models using the metrics viz., accuracy, F1-score, precision, and recall. The performance of the models was assessed using bias and variance. The best-fit models were evolved through trials, based on datasets splitting, learning rate, hyperparameter tuning, and optimizers. The analysis of the various CNN architectures and confirmation of their precision shall motivate the oncologist/physicians to use these intelligent models as an assistive tool for prognosis/diagnosis.