Nuclei instance segmentation is a crucial task in biomedical image analysis. However, nuclei instance segmentation is also a challenging task due to densely clustered nuclei and heterogeneity in the nuclei phenotype including size and shape. Numerous postprocessing methods suitable for deep-learning-based nuclei instance segmentation have been proposed. This project aims to compare the following postprocessing methods:
- A contour-based postprocessing method that is similar to the postprocessing methods of Chen et al. 2017 and Zhou et al. 2019
- The postprocessing method from “Nuclei segmentation using marker-controlled watershed, tracking using mean-shift, and Kalman filter in time-lapse microscopy” by Yang et al. 2006
- The postprocessing method from “Segmentation of nuclei in histopathology images by deep regression of the distance map” by Naylor et al. 2019
- The postprocessing method from “HoVer-Net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images” by Graham et al. 2019
- A modified version of the postprocessing method from Graham et al. 2019
The postprocessing methods are further compared with a baseline postprocessing method incapable of separating touching/overlapping nuclei to allow for a better performance evaluation of the postprocesing methods. The evaluation of the postprocessing methods is conducted using
- the ground truth representations as input to determine the upper performance limit of the postprocessing methods.
- a single and a dual decoder U-Net (Ronneberger et al. 2015) to establish a baseline performance.
- a REU-Net (Qin et al. 2022) as a state-of-the-art network.
The MoNuSeg dataset (Kumar et al. 2017) is used for all experiments.