I don't know: Double-strategies based active learning for mammographie mass classification
Jingyang Zhang, Dong Chen, Hongzhi Xie, Shuyang Zhang, Lixu Gu · 2017
Automatic classification of mammographie mass is an important yet challenging task. Despite the great success of active learning applied to reducing labeling cost, the traditional active learning methods are not suitable for the real-world mammographic mass classification. Because of the uncertain and insufficient knowledge, the radiologist only presents “I don't know” to some queried mammographic cases. To solve this problem, the radiologist's knowledge information is modeled via diverse density concept. Then a novel double-strategies based active learning framework, which is an adaptive combination of the radiologist's knowledge information model and the most uncertainty sampling strategy, together with mutual information based sampling strategy, is developed. Each queried instance is guaranteed not only with high efficiency for training an accurate classifier, but also with high probability of belonging to radiologist's certain knowledge. Experiments on digital database for screening mammography demonstrate that our approach can obtain a reliable classifier for mammographic mass with fewer querying operations compared to traditional active learning methods.