Automatic Filter of Normal Papanicolaou Smear Using Multi-instance Learning Algorithms

Papanicolaou smear is a common method to detect cervical cancer. Along with the increase demand of detection, the workload of clinical doctors increases significantly. In this paper, we try to screen out absolute normal cervical smear using machine learning algorithms with the help of computers. The clinical images are preprocessed to reduce noise. The unsupervised learning method is then adopted and morphological operation is conducted in sequence to extract the cell nucleus in all images. Afterward, the key features of each instance are extracted for learning. The image sets are trained and tested in the multi-instance learning (MIL) framework. The results show that our proposed method can achieve satisfactory performance. Therefore, our proposed method can be expected by clinical doctors for use in clinical papanicolaou smear reading in the future.