Active reducing classification error for CAD systems

We address the problem of reducing classification error in a CAD system: given a small set of training examples with multiple labels from annotators whose reliabilities are unknown, the objective is to learn an effective classification model with as few mistakes as possible on other unlabeled examples. The problem usually occurs in the situation that there are no labeled examples to be utilized as “golden-standard” for testing the classification model or annotators. We propose an active scheme of obtaining an accurate classifier for CAD systems, by reducing classification error from two aspects of example and label selection. In every step of the iterative process, the classifier can automatic submit the most helpful examples of all to the annotators who are most likely to provide correct labels. The proposed scheme has been tested on two breast cancer datasets. Experimental results show that the proposed algorithm can achieve better accuracy than other existing methods.