Deep-learning framework based on a large ultrasound image database to realize computer-aided diagnosis for liver and breast tumors

The quality and quantity of training data is vital for computer-aided diagnosis (CADx) based on deep learning. However, the biomedical industry lacks large database of ultrasound images. Therefore, The Japan Society of Ultrasonics in Medicine (JSUM) is currently constructing an ultrasound image database for liver tumors, breast tumors, and heart diseases. As of August 2021, the project has collected more than 140,000 ultrasound images and videos. This database contains ultrasound images, their corresponding labels, and annotation information. That is, the ultrasound image data contains information related to the size and location of the tumor. In this study, we developed a CADx to classify liver tumors and breast tumors by utilizing approximately 71,000 liver tumor and 14,000 breast tumor ultrasound images from the abovementioned database. We classified liver tumors into four classes: cysts, hemangiomas, hepatocellular carcinomas, and metastatic liver cancers. Similarly, we classified breast tumors into four classes: breast cancer, fibroadenoma, cysts, and others. We used a convolutional neural network based on VGG19 for these classifications, and evaluated the accuracy of each case unit by k-fold cross-validation, thereby achieving an accuracy of 91.1% and 85.2% for four-class classification of liver tumor and breast tumor, respectively. In addition, the accuracy, sensitivity, and specificity of the benign/malignant classification based on this result was, respectively, 94.3%, 82.8%, and 96.7% for liver tumors and 89.9%, 92.6%, and 86.6% for breast tumors. Furthermore, when compared with the results obtained in a previous study that utilized a small database, using a large database provided a higher accuracy for both liver and breast tumors.