Label-Free Nuclei Segmentation Using Intra-Image Self Similarity

In computational pathology, nuclei segmentation from histology images is a fundamental task. While deep learning based nuclei segmentation methods yield excellent results, they rely on a large amount of annotated images; however, annotating nuclei from histology images is tedious and time-consuming. To get rid of labeling burden completely, we propose a label-free approach for nuclei segmentation, motivated from one pronounced yet omitted property that characterizes histology images and nuclei: intra-image self similarity (IISS), that is, within an image, nuclei are similar in their shapes and appearances. First, we leverage traditional machine learning and image processing techniques to generate a pseudo segmentation map, whose connected components form candidate nuclei, both positive or negative. In particular, it is common that adjacent nuclei are merged into one candidate due to imperfect staining and imaging conditions, which violate the IISS property. Then, we filter the candidates based on a custom-designed index that roughly measures if a candidate contains multiple nuclei. The remaining candidates are used as pseudo labels, which we use to train a U-Net to discover the hierarchical features distinguish nuclei pixels from background. Finally, we apply the learned U-Net to produce final nuclei segmentation. We validate the proposed method on the public dataset MoNuSeg. Experimental results demonstrate the effectiveness of our design and, to the best of our knowledge, it achieves the state-of-the-art performances of label-free segmentation on the benchmark MoNuSeg dataset with a mean Dice score of 79.2%.