Background Subtraction Approach to Unsupervised Cell Segmentation: Toward Excluding Spurious Features in Degraded Cytology Slides

Automation and augmentation of the cytological diagnosis by data-driven techniques require a large number of microscopic slide images. The slide images with different imaging and storage conditions may contain latent image features that are spuriously correlated with cell classes. This paper focuses on the fact that image features other than cells can induce inference in deep learning of cytology images. To address this problem without the cost of cell annotation, this paper proposes a novel image processing method that subtracts from the whole image the background pattern discolored by the degradation of the slides. In the proposed method, a deep neural network model learns to separate sparse cells from the background in an unsupervised manner. To demonstrate the capability of the proposed method to exclude the background features, it is experimentally shown that it achieves cell detection with a comparable performance to that of supervised methods.