The Japan Society of Ultrasonics in Medicine (JSUM) is currently constructing an ultrasound image database. This database collects B-mode images of liver tumors and breast tumors, and B-mode videos of heart disease. In the past year, 31,000 liver tumor images have been collected from 11 institutions and 14,000 breast tumor images have been collected from 5 institutions. We are developing computer-aided detection (CADe) and computer-aided diagnosis (CADx) systems for liver and breast tumors based on deep learning using this database. In this paper, we report on CADx to estimate liver tumor types as a first trial. The data used in this study are 159 cyst cases (338 images), 68 hemangioma cases (279 images), 73 hepatocellular carcinoma (HCC) cases (241 images), and 24 metastatic liver cancer cases (122 images), collected at one facility. We developed the CADx system that estimates four types of liver tumor using a convolutional neural network based on VGGNet. The accuracy of the developed 4-class classification CADx was 88.0%. The accuracy by tumor type was 98.1% for cysts, 86.8% for hemangiomas, 86.3% for HCC, and 29.2% for metastatic liver cancer, with increasing accuracy observed for larger data sets. We also developed CADx to estimate whether a liver tumor is benign or malignant. The accuracy of this 2-class classification CADx was 94.8%, the sensitivity was 93.8%, and the specificity was 95.2%. Both 4-class classification and 2-class classification CADx had relatively high accuracy. However, in this study, we used only a small amount data collected from a single facility. In the future, we plan to verify our results using a larger amount of data collected from multiple facilities. In addition, we prototyped CAD software and are currently developing it with feedback from doctors.