Differentiation between Phyllodes Tumors and Fibroadenomas through Breast Ultrasound: Deep-Learning Model Outperforms Ultrasound Physicians

Abstract

The preoperative differentiation of breast phyllodes tumors (PTs) from fibroadenomas (FAs) plays a critical role in identifying an appropriate surgical treatment. Although several imaging modalities are available, reliable differentiation between PT and FA remains a great challenge for radiologists in clinical work. Artificial intelligence (AI)-assisted diagnosis has shown promise in distinguishing PT from FA. However, a very small sample size was adopted in previous studies. In this work, we retrospectively enrolled 656 breast tumors (372 FAs and 284 PTs) with 1945 ultrasound images in total. Two experienced ultrasound physicians independently evaluated the ultrasound images. Meanwhile, three deep-learning models (i.e., ResNet, VGG, and GoogLeNet) were applied to classify FAs and PTs. The robustness of the models was evaluated by fivefold cross validation. The performance of each model was assessed by using the receiver operating characteristic (ROC) curve. The area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were also calculated. Among the three models, the ResNet model yielded the highest AUC value, of 0.91, with an accuracy value of 95.3%, a sensitivity value of 96.2%, and a specificity value of 94.7% in the testing data set. In contrast, the two physicians yielded an average AUC value of 0.69, an accuracy value of 70.7%, a sensitivity value of 54.4%, and a specificity value of 53.2%. Our findings indicate that the diagnostic performance of deep learning is better than that of physicians in the distinction of PTs from FAs. This further suggests that AI is a valuable tool for aiding clinical diagnosis, thereby advancing precision therapy.