Objective:

To investigate the ipsilateral axillary lymph node metastases based on high-throughput sonographic features of invasive breast carcinoma.

Methods:

In order to represent the characteristics of breast tumors

the high-throughput features based on Breast Imaging Reporting and Data System (BI-RA

DS)

features based on scale-invariant feature transform (SIFT) and features based on convolutional neural network (CNN) were extracted. Combined with three important clinical features

a total of 10 703 features were obtained to constitute a feature system. A 10-fold feature selection method using t-test and least absolute shrinkage and selection operator (LASSO) regression model was bootstrapped for 100 times. The 100 most frequent features were input into a linear-kernel support vector machine (SVM) classifier in 10-fold cross validation experiment.

Results:

For the included 380 breast cancer cases

the area under curve (AUC) of the receiver operating characteristic (ROC) curve

accuracy

sensitivity

specificity were 0.903

82.6%

90.6%

69.9%

respectively.

Conclusion:

The sonographic features extracted from breast ultrasound images can predict lymph node metastases and provide the possibility of lymph node metastases for a sonographer.