A PSVM-Based Active Learning Method for Mass Detection

In mammograms, masses always vary widely in their shapes and densities, and yet share common appearances with the normal tissues. This point extremely increases the detection difficulty and also impacts the performance of the automatic mass detecting system. To improve the sensitivity of mass detection system, we propose an active learning scheme to detect various masses on mammograms. Firstly, the pairwise constraints are introduced, and the scheme conducts with pairwise support vector machine (PSVM) by involving the relationship among different samples into the classification procedure. Furthermore, according to the detection results, the missed samples with their uncertainty information are combined with the matched feature distance among different samples to provide for re-consideration. Then, with the representative information, the proposed PSVM-based method actively selects the pairwise samples that should be feed back to the training set. The experimental results show that the proposed active learning method with PSVM could make full use of the information of samples, and thus, it could get satisfactory detection rates and false positives during the detection procedure. The method can also get good compromise between the sensitivity and specificity, and the whole learning scheme has better generalization ability and detection performance in comparison with some existing detection methods.