ABUS-Net: Graph convolutional network with multi-scale features for breast cancer diagnosis using automated breast ultrasound

Abstract

Breast cancer is the leading cause of cancer-related deaths in women. Early screening helps with the treatment and recovery of breast cancer. The automated breast ultrasound (ABUS), as a standardized 3D breast ultrasound imaging technology, overcomes the limitations of traditional ultrasound, such as strong dependence on operator skills and poor reproducibility. At the same time, the coronal view provided by ABUS contains rich information that aids in the diagnosis of breast cancer. Therefore, how to effectively utilize the coronal features and spatial structural relationships is an urgent challenge. To address this issue, a graph convolutional network (GCN) with multi-scale features is proposed for breast cancer diagnosis using ABUS, named ABUS-Net. Unlike previous studies that relied on 3D patch techniques to represent tumor spatial features, our method focuses on the deep exploration of coronal features while using a GCN model to capture the spatial structural relationships of breast tumors. Specifically, we design a multi-scale feature extraction module to capture detailed information at different scales in the ABUS coronal sections, thereby enhancing the tumor feature representation. We then treat each slice containing tumor as a graph vertex, with the inherent spatial relationships between slices forming the edges. Finally, we employ the GCN model to classify the malignancy of the breast tumor. To validate the effectiveness and superiority of the model, we test it on both private and public datasets and compare it with other existing models. Experimental results highlight the potential utility of the proposed model in clinical practice.