Mining diagnostic rules of breast tumor on ultrasound image using cost-sensitive RuleFit method

In the medical diagnosis, the false negative prediction is more serious than the false positive prediction. We introduce the cost-sensitive rule ensemble method (RuleFit) to breast ultrasound, which can induce the interpretable scoring rules for malignancy assessment, and can be applied to tune the sensitivity and specificity of the predictive model by varying the cost weights of misclassification. The GentleCost boosting algorithm is proposed to generate the decision tree ensemble. Then, we use the modified RuleFit method with the cost-weighted loss function to select and fit the rules decomposing from the tree ensemble. Experiments results on a breast ultrasound image dataset (168 cases) with the varying cost weights demonstrate that the final rule ensemble contain only 22 (among total 600 decomposed rules) rules with the comparable performance to the tree ensemble. The examples of the rule ensemble for breast ultrasound and its interpretation are also illustrated.