Filtering normal papanicolaou smear with multi-instance learning

Abstract

Filtering normal papanicolaou smear using computer-aided system can help clinical doctors to detect cervical cancer. In this paper, we propose a scheme to classify cervical cells as normal or abnormal. The pipeline includes preprocessing, perinuclear area extraction, feature extraction and multi-instance learning (MIL). We tried and compared several feature extraction methods, including textural features, manual features and Stacked sparse autoencoder(SSAE) self-learned features. In multi-instance learning, we modify softmax classifier to be adequate for our problem besides some classic MIL algorithms. The results show that manual features or SSAE with modified softmax achieve the best performance and are recognized by clinical pathology doctors.