Computer-aided diagnosis system for breast ultrasound images using deep learning

Abstract

Introduction Ultrasonography has been recommended as an adjunctive modality to mammography, which is insufficient for accurate diagnosis of breast cancer in women with dense breast tissue. However, ultrasonography has the disadvantage of being operator dependent, requiring proficiency in reading ultrasound (US) images and increasing the false positive rate. To overcome these problems, we aimed to develop a computer-aided diagnosis (CAD) system for classification of breast malignant and benign masses on ultrasonography based on the convolutional neural network (CNN), a state-of-the-art deep learning technique. Methods From a large ultrasound (US) image database, managed by Japan Association of Breast and Thyroid Sonology (JABTS), we collected images of 1536 breast masses (897 malignant and 639 benign) confirmed by pathological examinations, with each breast mass captured from various angles using the US imaging probe. We constructed an ensemble network by combining two CNN models (VGG192 and ResNe1523) fine-tuned on balanced training data with the data augmentation which is a popular technique for synthetically generating new samples from the original and used the mass level classification method enabling the CNN to classify a given mass using all views. We explored the grounds of classification by generating a heatmap capable of presenting important regions used by the CNN for classification in humans. Results On an independent 154 test masses (77 malignant and 77 benign), our network showed outstanding classification performance with a sensitivity of 90.9%, a specificity of 87.0%, and an AUC of 0.951 compared with the two CNN models. In addition, our study indicated that not only breast masses but also surrounding tissues are important regions for correct classification. Conclusion This CNN-based CAD system is expected to assist doctors from the viewpoint of second opinion and improve the diagnosis of breast cancer in clinical practice.