The segmentation of nuclei from histopathology images with synthetic data
Abstract
The segmentation of nuclei from histopathology images provides the necessary information to diagnose cancer. Existing deep learning-based nuclei segmentation approaches demand a large amount of annotated data from expert pathologists where the size of annotated datasets is relatively small. This paper proposes the generation of synthetic annotated data utilising a small number of labelled data. The background of synthetic patches is created using base image patches, where the generative adversarial model is applied to preserve the characteristics of texture regions. By deploying a feature learning-based CNN filtering method, more realistic backgrounds are assigned to the highest probability, and the backgrounds with lower probability scores are discarded. The foreground nuclei shapes are collected from the original images using ground truth masks, and the shape transformations are applied using image processing techniques. These nuclei shapes are placed on the filtered background by preserving the characteristics of shape and distribution of the nuclei. The synthetic image patches with corresponding masks are used to train the modified U-net segmentation network in order to segment the nuclei regions. The segmentation results in terms of DSC and JI are 0.875 and 0.791, respectively. Synthetic images performed better than original images across a range of diverse histopathology image backgrounds.
RAS ID
58218
Document Type
Journal Article
Date of Publication
2023
School
School of Science
Copyright
subscription content
Publisher
Springer
Comments
Hossain, M. S., Armstrong, L. J., Abu-Khalaf, J., & Cook, D. (2023). The segmentation of nuclei from histopathology images with synthetic data. Signal, Image and Video Processing, 17, 3703–3711. https://doi.org/10.1007/s11760-023-02597-w