Exploring Residual Networks for Breast Cancer Detection from Ultrasound Images

Arijit Das, Srinibas Rana

Abstract: Breast cancer is one of the most leading causes of cancer globally, affecting mostly women at any age after their puberty. Although the disease is fatal and leads to thousands of deaths yearly, it is mostly curable if detected in its early stages. As a result, fast and accurate detection techniques play a crucial role in patient survival. Earlier, doctors resorted to manual detection techniques for this purpose. However, these techniques were slow and often subjective to the expertise of the doctor. Thus, with the advancement of technology, these primitive methods were replaced by computer-aided detection algorithms. One such algorithm is the convolutional neural network (CNN). It is a deep learning algorithm that has demonstrated promising results in medical image analysis in the past few years. It possesses special neural receptors called filters, which can automatically extract features from an image without human intervention. This fascinating property has urged several researchers to propose different CNN architectures in the past few years, each having its unique methodology. In this work, the residual network (ResNet) architecture is studied for detecting breast cancer from ultrasound images. ResNets have special skip connections which make them superior to other architectures. Depending on the depth, a ResNet can be categorized into different types - ResNet-18, ResNet-34, ResNet-50, ResNet-101, and ResNet-152 being some widely used ones. In this research, each of these five ResNets is trained on the Breast Ultrasound Images (BUSI) dataset twice - first from scratch and second using transfer learning. Thereafter, their performances are evaluated and compared using four metrics - F1-score, precision, recall, and accuracy. The confusion matrices are also plotted for a better understanding of the results. Lastly, the class activation heatmaps are generated using GradCAM++ to visualize the crucial regions of the ultrasound images as interpreted by the ResNet architectures. The code used to conduct this research can be accessed at https://github.com/iamarijit/US-cancer-resnet.