A Multi-scale Self-supervision Method for Improving Cell Nuclei Segmentation in Pathological Tissues

Abstract

AbstractNuclei detection and segmentation in histopathological images is a prerequisite step for quantitative analysis including morphological shape and size to help in identifying cancer prognosis. Digital pathology field aims to improve the quality of cancer diagnosis and has helped pathologists to reduce their efforts and time. Different deep learning architectures are widely used recently in Digital pathology field, yielding promising results in different problems. However, Deep convolutional neural networks (CNNs) need a large subset of labelled data that are not easily available all the time in the field of digital pathology. On the other hand, self-supervision methods are frequently used in different problems with the aim to overcome the lack of labelled data. In this study, we examine the impact of using self-supervision approaches on the segmentation problem. Also, we introduce a new multi-scale self-supervision method based on the zooming factor of the tissue. We compare the proposed method to the basic segmentation method and other popular self-supervision approaches that are used in other applications. The proposed Multi-scale self-supervision approach is applied on two publicly available pathology datasets. The results showed that the proposed approach outperforms Baseline U-Net by 0.2% and 0.02% for nuclei segmentation–mean Aggregated Jaccard Index (AJI), in TNBC and MoNuSeg, respectively.KeywordsNuclei segmentationHistopathological imagesSelf-supervisionTransfer learningDeep learning