Application of deep learning to establish a diagnostic model of breast lesions using two-dimensional grayscale ultrasound imaging.

PURPOSE There are currently few specific artificial intelligence (AI) studies for Breast Imaging Reporting and Data System (BI-RADS) category 4A lesions. This study aimed to establish an AI diagnostic model of breast lesions using two-dimensional grayscale ultrasound imaging and to compare its performance with that of radiologists. METHODS The ultrasound images of 1311 lesions were evaluated by radiologists according to the BI-RADS categories, using pathology results as reference. Two classification standards (standards 1 and 2) for benign and malignant lesions were defined and used to calculate the diagnostic performance of radiologists, altogether and individually. The breast lesion images were also used to develop an AI diagnostic model. RESULTS The diagnostic performance of AI and that of the radiologists were compared using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). All parameters of diagnostic performance, except for sensitivity and NPV, improved with standard 2. For the 202 lesions in the test set, the diagnostic performance of the AI model had 77.0% accuracy, 82.0% sensitivity, 71.7% specificity, 79.3% PPV, 75.1% NPV, and an AUC of 0.846. When the AI model was used to analyze category 4A lesions, the PPV was 9.3%, which was better than that of the radiologists, although not significantly. CONCLUSIONS Deep learning technology shows a good performance in classifying benign and malignant breast lesions. It may be potentially used in practice to improve diagnostic accuracy and reduce unnecessary biopsies of breast lesions.