Computer-aided diagnosis of breast cancer in ultrasonography images by deep learning

• Published in: Neurocomputing (Amsterdam) Vol. 472; pp. 152 - 165
• Main Authors: Qi, Xiaofeng, Yi, Fasheng, Zhang, Lei, Chen, Yao, Pi, Yong, Chen, Yuanyuan, Guo, Jixiang, Wang, Jianyong, Guo, Quan, Li, Jilan, Chen, Yi, Lv, Qing, Yi, Zhang
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-02-2022
• Subjects: Breast cancer; Computer-aided diagnosis; Deep neural networks; Initial screening; Ultrasonography; Breast cancer; Deep neural networks; Initial screening; Computer-aided diagnosis; Ultrasonography
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2021.11.047

Ultrasonography of the breast mass is an important imaging technology for diagnosing breast cancer. In China, ultrasound equipment is widely used in medical institutions. Patients obtain a report with highlighted ultrasonography images through an initial clinical screening. However, analyzing these images manually is highly subjective for the variation in the clinical competence of doctors, resulting in poor consistency and low sensitivity. In this study, an automated breast cancer diagnosis system is developed to increase diagnostic accuracy. The system is deployed on mobile phones, takes a photo of the ultrasound report as input and performs diagnosis on each image. The developed system consists of three subsystems. The first subsystem is to reduce noise in the taken photos, reconstructing high-quality images. We develop the first subsystem based on the frameworks of stacked denoising autoencoders and generative adversarial networks. The second subsystem is to classify images into malignant and non-malignant; to extract high-level features from the images, deep convolutional neural networks are employed. The third subsystem is to detect anomalies in model performances, reducing false negative rates. Generative adversarial networks are utilized to distinguish false negative samples from true negative samples. 18,225 breast ultrasonography images and 2416 ultrasound reports are collected to train and evaluate the system. Experimental results show that the performance of our system is comparable to that of human experts. It is believed that this is the first system for breast cancer diagnosis deployed on mobile phones. The developed system is integrated with a cloud computing platform and accessible online to aid in the initial screening and diagnosis of breast cancer, thereby promoting earlier treatment, reducing the morbidity and mortality.