Lizard: A Large-Scale Dataset for Colonic Nuclear Instance Segmentation and Classification

Abstract

Nuclear detection, segmentation and morphometric profiling are essential in helping us further understand the relationship between histology and patient outcome. To drive innovation in this area, we setup a community-wide challenge using the largest available dataset of its kind to assess nuclear segmentation and cellular composition. Our challenge, named CoNIC, stimulated the development of reproducible algorithms for cellular recognition with real-time result inspection on public leaderboards. We conducted an extensive post-challenge analysis based on the top-performing models using 1,658 whole-slide images of colon tissue. With around 700 million detected nuclei per model, associated features were used for dysplasia grading and survival analysis, where we demonstrated that the challenge's improvement over the previous state-ofthe-art led to significant boosts in downstream performance. Our findings also suggest that eosinophils and neutrophils play an important role in the tumour microevironment. We release challenge models and WSIlevel results to foster the development of further methods for biomarker discovery.

Panoptic segmentation with highly imbalanced semantic labels

“…Yet, further work in this domain is required as we still observe a number of misclassifications as well as nuclei not being detected correctly. Moreover, the Lizard dataset [5] aggregated some cell types (e.g. muscle, fibroblasts, endothelial cells) and excluded a number of other highly relevant cells such as macrophages and dendritic cells as well as basophils and mast cells.…”

“…The challenge provides the previously published Lizard Dataset [5] and required participants to exclusively make use of this dataset. The challenge organizers also provided a pre-tiled version of the same dataset which we used for all experiments.…”

“…The challenge dataset of H&E stained tissue tiles comes with drastic semantic class imbalance: Among the six cell types described in [5], neutrophils and eosinophils each make up less than 1% of the cells in the dataset. Furthermore, nucleus size varies considerably across cell types, while at the same time some cell types, though frequent, also exhibit * equal contribution high intra-class variation in shape (see Section III-A).…”

Class-Controlled Copy-Paste Based Cell Segmentation for CoNIC Challenge

“…We follow the HoverNet [2] but with some differences. We replace the HoverNet [2] model structure with Unet++ [3] and modify the structure so that model can output 3 outputs like HoverNet [2]. We select EfficientNet-b7 [4] pretrained on ImageNet [5] for encoder of Unet++ [3].…”

“…This manuscript describes submitted method of team Arontier to CoNIC Challenge [1][2]. Our method mainly focus on how to deal with the class imbalance using copy-paste augmentation [1] with a little bit modification.…”