Semi-Supervised Cell Instance Segmentation for Multi-Modality Microscope Images

Wang Ziyue, Zijie Fang, Yang Chen, ZeXi Yang, Xinhao Liu, Yongbing Zhang

Keywords: cell segmentation, multi-modality, semi-supervised learning, microscope images

TL;DR: This paper proposes an ensemble model which fulfills semi-supervised cell segmentation for microscope images with multi-modality.

Abstract: Many clinical and biological tasks depend on accurate cell instance segmentation. Currently, the rapid development of deep learning realizes the automation of cell segmentation, which significantly decreases the workload of clinicians and researchers. However, most existing cell segmentation frameworks are fully supervised and modality-specific. Towards this end, this paper proposes a semi-supervised cell instance segmentation framework for multi-modality microscope images. Firstly, $K$-Means clustering is utilized to discriminate the image modality. Then, for phase contrast and differential interference contrast images, Cellpose is adopted. For brightfield images, we subdivide them into two sub-categories according to the cell diameter by $K$-Means and optimize a U-Net for the large diameter group. For fluorescence images, we propose a semi-supervised learning strategy using CDNet. The leaderboard shows that our proposed framework reaches an F1 score of 0.8428 on the tuning set, which ranks 6th among all teams.