SISU: A Holistic Self-training Framework on Semi-supervised White Blood Cell Segmentation

Quang-Hien Kha, Minh Huu Nhat Le, Lam Huu Phuc Nguyen, Minh Nguyen Tuan Tran, Linh My Nguyen, Hung Quay Thong, Nguyen-Quoc-Khanh Le

Abstract: The automated segmentation of leukocytes plays a vital role in diagnosing and monitoring life-threatening conditions such as leukemia and lymphoma. However, this task encounters challenges due to the limited availability and quality of public datasets. To effectively utilize the limited dataset, it is necessary to develop a semi-supervised learning framework. In this regard, we propose a holistic self-training framework called Self-training using In-turn Supervised-Unsupervised training (SISU). Our framework encompasses two key components. Firstly, we introduce Feature Perturbed Cross-View Co-Training, which incorporates dual feature perturbation methods and utilizes two auxiliary decoders to enhance the robustness of the feature representation. Secondly, drawing inspiration from FixMatch, we integrate a regularization mechanism via weak-to-strong consistency training to further enhance the self-training framework. Finally, we conduct training and evaluation of SISU for semi-supervised learning on three datasets (Zheng 1, Zheng 2, and LISC), achieving remarkable mIOU scores of up to 93.54%, 88.92%, and 77.02% respectively across multiple settings within the self-training scheme.