Medical Knowledge Guided Intra-Specimen Reference Network for Cervical Cell Classification.

Deep learning-based method for automatic identification of cervical cells has become a research hotspot in recent years. However, few studies have explored effective ways to incorporate medical domain knowledge. In this paper, we propose a novel deep learning network named intra-specimen reference network (IsrNet) for automatic identification of cervical cells, which models the relationship of different cervical cells and extracts contrastive information from the same specimen. Compared to the general classification process, IsrNet reconstructs the input format and leverages a joint loss function. In addition, we utilize weighted information entropy (WIE) to measure the diversity degree of a specific cervical cell dataset. We collect a cervical cell dataset, LandingTBS, and build four new datasets based on it to acquire data with different WIE. The experimental results demonstrate that the intra-specimen reference feature is effective and significant for cervical cell classification. Besides, IsrNet is able to extract more contrastive information among different cells from the same specimen by using the dataset with high WIE. Via mimicking the way a cytopathologist diagnoses cervical cells, IsrNet provides a simple and effective approach for high-performance identification of cervical cells.