Computerized Classification Method for Histological Classification of Masses on Breast Ultrasonographic Images Using Convolutional Neural Networks with ROI Pooling

Abstract

It can be difficult for clinicians to correctly determine histological classifications of masses on breast ultrasonographic images. The purpose of this study was to develop a computerized classification method for histological classification of masses on breast ultrasonographic images using convolutional neural networks (CNN) with a ROI pooling that analyzes feature maps focusing on the mass region. Our dataset consisted of 585 breast ultrasonographic images obtained from 585 patients. It included 288 malignant masses (218 invasive and 70 noninvasive carcinomas) and 297 benign masses (115 cysts and 182 fibroadenomas). In this study, we developed a modified CNN model based on ResNet-18, in which the ROI pooling and two fully connected layers with a softmax function were introduced after the second and fourth residual block on ResNet-18, respectively. The proposed CNN model was employed to distinguish among four different types of histological classifications for masses. A three-fold cross validation method was used for training and testing the proposed CNN model. The average accuracy, sensitivity, specificity, positive predictive value and negative predictive value for the proposed CNN model were 81.7%, 91.0%, 91.2%, 91.0%, and 91.2%, respectively. Those results were substantially greater than those with ResNet-18 (70.3%, 83.0%, 87.2%, 86.3%, and 84.1%).