PROJECT TITLE :

A Multi-Organ Nucleus Segmentation Challenge

ABSTRACT:

The development and validation of visual biomarkers for new digital pathology datasets can considerably benefit from the use of generalised nucleus segmentation algorithms. For the MoNuSeg 2018 Challenge, the goal was to develop nuclei segmentation algorithms that could be used in all digital pathology workflows. More than 80 people from 32 different institutions took part in this official satellite event of the MICCAI 2018 conference. There were 30 photos from seven organs with annotations on 21,623 nuclei presented to the contestants as a training set. Only 14 photos from seven different organs, including two that weren't in the training set, were released for testing. Prioritizing instance segmentation over Semantic Segmentation, entries were ranked using the average aggregated Jaccard index (AJI) on the test set. It was found that more than half of the teams that completed the challenge exceeded a previous baseline. Color normalisation and substantial data augmentation were among the trends that contributed to enhanced accuracy. U-Net, FCN, and Mask RCNN inspired fully convolutional networks were also popularly employed, often based on ResNet or VGG base designs. On anticipated semantic segmentation maps, a common post-processing technique was called "watershed" segmentation. In comparison to a single human annotator, several of the top approaches performed admirably, and as a result, they can be employed confidently in nuclear morphometrics.