MT2: Multi-task Mean Teacher for Semi-Supervised Cell Segmentation

Cell segmentation is significant for downstream single-cell analysis in biological and biomedical research. Recently, image segmentation methods based on supervised learning have achieved promising results. However, most of them rely on intensive manual annotations, which are extremely time-consuming and expensive for cell segmentation. In addition, existing methods are often trained for a specific modality with poor generalization ability. In this paper, a novel semi-supervised cell segmentation method is proposed to segment microscopy images from multiple modalities. Specifically, Mean Teacher model is introduced to a multi-task learning framework, named Multi-task Mean Teacher (MT2), in which both the classification and the regression heads are utilized to improve the prediction performance. Moreover, new data augmentation and multi-scale inference strategies are presented to enhance the robustness and generalization ability. For the quantitative evaluation on the Tuning Set of NeurIPS 2022 Cell Segmentation Competition, our method achieves the F1 Score of 0.8690, which demonstrates the effectiveness of the proposed semi-supervised learning method. Code is available at https://github.com/djh-dzxw/MT2.