Unsupervised Cell Localization and Segmentation via Synchronized Spectral Clustering

In this paper, we propose a unified model, deep synchronized spectral clustering (DeepSSC), for unsupervised cellular structure detection and segmentation. Unlike previous deep spectral clustering methods, which handle segmentation of single or small sets of salient objects through iterative mask-cuts and additional synchronization in postprocessing to correct cross-image clustering inconsistencies, DeepSSC constructs a joint spectral embedding and handles synchronized cluster assignments by capturing both inter-and intra-image relevance to identify relevant cellular structures. Additionally, we fine-tune the pre-trained model using self-supervised learning to address domain shifts caused by varying appearances and textures in microscopic images. DeepSSC is designed to simultaneously detect and segment large numbers of fine-grained, repetitively occurring cells. Experimental results demonstrate that DeepSSC achieves robust, superior performance compared to state-of-the-art methods.