A Performance Evaluation of Machine Learning Techniques for Breast Ultrasound Classification

In this paper, a performance comparison of seven machine learning (ML) approaches for classifying breast lesions on ultrasound is presented. From a dataset with 2032 cases (1341 benign and 691 malignant), 137 morphological and texture features were extracted, aiming to describe the BI-RADS lexicon for masses. Support vector machine (SVM), k -nearest neighbor ( kNN), radial basis function network (RBFN), linear discriminant analysis (LDA), multinomial logistic regression (MLR), random forest (RF), and AdaBoost (Ada) were evaluated in terms of sensitivity (SEN), specificity (SPE), accuracy (ACC), and area under the ROC curve (AUC). The results revealed that LDA obtained the best classification performance with ACC=0.89,SEN=0.82, SPE=0.93, and AUC=0.95. Contrarily, the kNN obtained the lowest classification performance with ACC=0.87, SEN=0.76, SPE=0.93 and AUC=0.91. This results point out that the LDA classifier can be convenient to be used in a CAD system because it is simple to implement, and it does not require the tuning of hyperparameters.