Deep Learning for Automated Breast Cancer Detection in Ultrasound: A Comparative Study of Four CNN Architectures

Breast cancer is one of the most common malignancies among women globally, and it constitutes a significant publichealth problem in terms of morbidity and mortality. Since early-stage diagnosis significantly increases treatment success and survival rates,effective screening and diagnostic methods are of great importance. Various imaging modalities, such as mammography, ultrasonography(US), and magnetic resonance imaging, play a critical role in the detection of breast cancer. Ultrasound, in particular, is a valuable imagingmethod due to its non-ionizing nature, its accessibility, and its role as a complementary tool in dense breast tissue. In recent years,deep learning (DL) algorithms, particularly Convolutional Neural Networks (CNNs), have exhibited promising results in medical imageanalysis, especially in cancer detection. The aim of this research is to investigate and compare the four most common CNN architectures,ResNet50, DenseNet169, InceptionV3 and InceptionV4, for breast ultrasound images to classify breast cancer automatically. We haveutilized publicly available breast ultrasound image datasets for the models and reported results in metrics of accuracy, precision, sensitivity,and F1-score. The InceptionV3 architecture had the best performance across the models examined with metrics of accuracy: 96.67%,precision: 96.55%, sensitivity: 96.38%, and F1-score: 96.41%. It was also noticed that the DenseNet169 model performed similarly to theInceptionV3 model but had substantially fewer parameters. The results of this study suggest that the InceptionV3 DL architecture mayhave significant potential for accuracy in the classification of cancer from breast ultrasound images and can contribute to the developmentof computer aided diagnosis systems for the early detection of breast cancer.