Cell Segmentation in Multi-modality High-Resolution Microscopy Images with Cellpose

Hyungjo Byun, Kwanyoung Lee, Hyunjung Shim

Keywords: Deep Learning, Instance Segmentation, Cell Segmentation

TL;DR: We made a cell segmentation model that applied cellpose model for "Weakly Supervised Cell Segmentation in Multi-modality High-Resolution Microscopy Images" challenge

Abstract: Deep learning has achieved great results in microscopy image processing in the field of Biology. However, a generalized model is needed which can solve overfitting and produce good performance for test images and unseen classes. The reason why it is difficult to make a generalized model is because of the diversity of modality, staining methods, cell shapes and resolution of microscopy images. The dataset of the "Weakly Supervised Cell Segmentation in Multi-modality High-Resolution Microscopy Images" challenge consists of images with these diverse characteristics. Therefore, we trained cellpose to perform instance segmentation well on dataset having various characteristics. In order to apply the cellpose model to the challenge dataset, we designated model to always use green and blue channels for any type of images. What’s more we newly created the size estimation model predicting diameter of cell to operate on various resolutions. As a result, we could achieve F1 score 0.7607 for the validation (Tuning) set.