Detection of Breast Masses in Mammograms by Incremental Discriminant Based Support Vector Machine Classifier and Active User Involvement

Abstract

Breast masses are among the most studied mammary pathologies on mammographic images. The complexity and diversity of mass forms require the use of appropriate descriptors and techniques. In this work, we propose a mass detection process based on our novel incremental Discriminant Based Support Vector Machine classifier coupled with the active involvement of domain experts. It is a three steps process. In the first step, mammographic images are pre-processed by eliminating noise and enhancing contrast. The second step is a feature engineering one. Multiple descriptors are extracted from the mammographic images and feature space transformation and reduction are performed by PCA (Principal Component Analysis). In the third step, the mass detection is performed based on the incremental classification guided by user involvement. The user is first provided with a simple user interface. Hence, this interface allows him to interact with the mammographic image and select some parts. Based on the selection, the pixels are incrementally classified using the IDSVM, which combines both local and near-global variational information of the training data into the input space. Experiments on mammograms from the INbreast database, by reference to the ground truth results, demonstrate the validity of our method.