Deep Learning for Differentiation of Breast Masses Detected by Screening Ultrasound Elastography

Recently, deep learning using convolutional neural networks (CNNs) has yielded consistent results in image-pattern recognition. This study was aimed at investigating the effectiveness of deep learning using CNNs to differentiate benign and malignant breast masses identified by elastography on ultrasound screening. A data set of the elastography images of 245 breast masses (146 benign, 99 malignant) in 239 consecutive patients was retrospectively obtained. The data set was randomly split into training (55%), validation (25%) and test (20%) cohorts. A deep learning model predicting the probability of malignancy was constructed using GoogLeNet architectures (pre-trained by ImageNet) with 50 epochs. The model was then applied to the test data, and the results were compared with those obtained by evaluating the fat-to-lesion ratio (FLR) and by a 5-point visual color assessment (elasticity score). The receiver operating characteristic (ROC) curve was calculated to evaluate the performance of the model. The DeLong test was used to compare the areas under the ROC curve (AUCs). The CNN, FLR and elasticity score had a sensitivity of 0.800, 0.800 and 0.350; specificity of 0.966, 0.586 and 0.931; accuracy of 0.898, 0.673 and 0.694; positive predictive value of 0.941, 0.571 and 0.778; negative predictive value of 0.875, 0.810 and 0.675; and AUC of 0.895, 0.693 and 0.641, respectively. The AUC of the CNN was significantly higher than that of the FLR or elasticity score (p < 0.001). A CNN-based deep learning model for predicting benign or malignant breast masses revealed better diagnostic performance than did FLR or elasticity score-based estimations on ultrasound elastography. The CNN-based model also increased the positive predictive value from 57%-78% to 94%. Therefore, this model may reduce unnecessary biopsy recommendations for masses detected on breast ultrasound screening.