This study compares the performance of two proven but very different machine learners, Naïve Bayes and logistic
regression, for differentiating malignant and benign breast masses using ultrasound imaging.
Ultrasound images of 266 masses were analyzed quantitatively for shape, echogenicity, margin characteristics, and
texture features. These features along with patient age, race, and mammographic BI-RADS category were used to train
Naïve Bayes and logistic regression classifiers to diagnose lesions as malignant or benign. ROC analysis was performed
using all of the features and using only a subset that maximized information gain. Performance was determined by the
area under the ROC curve, Az, obtained from leave-one-out cross validation.
Naïve Bayes showed significant variation (Az 0.733 ± 0.035 to 0.840 ± 0.029, P < 0.002) with the choice of features, but
the performance of logistic regression was relatively unchanged under feature selection (Az 0.839 ± 0.029 to 0.859 ±
0.028, P = 0.605). Out of 34 features, a subset of 6 gave the highest information gain: brightness difference, margin
sharpness, depth-to-width, mammographic BI-RADs, age, and race. The probabilities of malignancy determined by
Naïve Bayes and logistic regression after feature selection showed significant correlation (R2= 0.87, P < 0.0001).
The diagnostic performance of Naïve Bayes and logistic regression can be comparable, but logistic regression is more
robust. Since probability of malignancy cannot be measured directly, high correlation between the probabilities derived
from two basic but dissimilar models increases confidence in the predictive power of machine learning models for
characterizing solid breast masses on ultrasound.