Color Space-based Hover-Net for Nuclei Instance Segmentation and Classification

Nuclei (grain) segmentation and classification is the first and most crucial step utilized in many different microscopic medical image analysis applications. However, from a deep-learning perspective, it proves to be a non-trivial task and suffers from many problems such as the presence of small objects, class imbalance, and having fine-grained differences between the different types of nuclei. In this paper, we introduce several contributions to tackle these problems. Firstly, we use the recently released ConvNeXt to replace the backbone of HoVer-Net to leverage the key components of transformers in a convolutional neural network. Secondly, we use a multi-channel color space-based approach to enhance the visual differences between nuclei to aid the model in extracting distinguishing features. Thirdly, we use a Unified Focal loss (UFL) to tackle the problem of background-foreground imbalance. Finally, we use a Sharpness-Aware Minimization (SAM) optimizer to improve the generalizability of the model. Overall, we are able to outperform the current state-of-the-art model, HoVer-Net, on the preliminary test set of the CoNiC Challenge 2022 by 12.489% mPQ+.