Fully Automated Interpretable Breast Ultrasound Assisted Diagnosis System

Breast cancer is one of the most vulnerable malignant tumors for women in the world, which seriously threatens women's life and health. Fortunately, through timely screening, early breast cancer patients have a high cure rate, which is expected to reach complete recovery. Ultrasound imaging is a screening technology for breast cancer that is superior to the pathological judgment of puncture. It is less harmful to the patient and has the advantages of easy operation and real-time, so it is currently widely used. In order to alleviate the shortage of medical resources and improve screening efficiency, the technology of the breast cancer computer-aided diagnosis system came into being. This paper proposes a new and interpretable fully automated breast ultrasound computer-aided diagnosis method, which can automatically extract the corresponding medical features used by doctors after inputting an ultrasound image and infer the final diagnosis result through the knowledge association between these features. First, the input ultrasound image is read intelligently, similar to the reading process of ultrasound doctors. The diagnostic system extracts medical features useful for diagnosis from the ultrasound image. The second step is associating the extracted features with the known breast benign and malignant diagnosis knowledge graph and finally determining the benign and malignant tumor in this ultrasound image through the correlation relationship in the knowledge graph which is presented as a knowledge tensor. The experimental results show that the fully automated breast ultrasound CAD system proposed in this paper can realize the tradeoff between interpretability and accuracy.