Diagnosing breast tumors from MRI appearances: cross-validation of neural network and logistic regression models

An algorithmic model based on the logistic regression analysis and a non-algorithmic model based on the artificial neural network (ANN) were established. The ability of these models was compared with each other in clinical application to differentiate malignant from benign breast tumors in a study group of 161 patients' records. Each patient’s r -cord consisted of 6 subjective features extracted from MRI appearance. These findings were encoded as features for an ANN as well as a logistic regression model (LRM) to predict the outcome of biopsy. After both models had been trained perfectly on training samples (n=100); the validation samples (n=61) was presented to the trained network as well as the established LRMs. Finally, the diagnostic performance of models was compared to that of the radiologist in terms of sensitivity, pecificity and accuracy using receiver operating characteristic curve (ROC) analysis. The average output of the ANN yielded a perfect sensitivity (98%) and high accuracy (90%) similar to that of expert radiologist (96% and 92%); while specificity was smaller than the radiologist (67% verses 80%). The output of the LRM using significant features showed improvement in the specificity from 60% for the LRM using all features to 93% for the reduced logistic regression model while keeping the accuracy around 90%.

Keywords:

breast cancer , neural networks , logistic regression model , ROC curves