Semi-Supervised Semantic Segmentation using Redesigned Self-Training for White Blood Cells
Vinh Quoc Luu, Duy Khanh Le, Huy Thanh Nguyen, Minh Thanh Nguyen, Thinh Tien Nguyen, Vinh Quang Dinh
TL;DR
The paper tackles the lack of large labeled white blood cell (WBC) segmentation datasets by proposing a semi-supervised self-training framework that integrates FixMatch for consistency regularization. It adopts a two-stage ST/ST++ pipeline that leverages unlabeled data through pseudo-labeling and selective re-training, enabling end-to-end learning with both supervised and unsupervised losses. On Zheng1, Zheng2, and LISC datasets, the approach achieves its best performance with DeepLab-V3+ and ResNet-50, achieving $90.69\%$, $87.37\%$, and $76.49\%$ respectively. However, applying FixMatch during the supervised stage can reduce labeled-data accuracy, and pseudo-masks are less reliable on complex WBC images, indicating a need for domain-specific refinements. Overall, the method demonstrates the potential of semi-supervised learning for WBC segmentation and shows generalizability across datasets, albeit with limitations tied to intra- and inter-image variability.
Abstract
Artificial Intelligence (AI) in healthcare, especially in white blood cell cancer diagnosis, is hindered by two primary challenges: the lack of large-scale labeled datasets for white blood cell (WBC) segmentation and outdated segmentation methods. These challenges inhibit the development of more accurate and modern techniques to diagnose cancer relating to white blood cells. To address the first challenge, a semi-supervised learning framework should be devised to efficiently capitalize on the scarcity of the dataset available. In this work, we address this issue by proposing a novel self-training pipeline with the incorporation of FixMatch. Self-training is a technique that utilizes the model trained on labeled data to generate pseudo-labels for the unlabeled data and then re-train on both of them. FixMatch is a consistency-regularization algorithm to enforce the model's robustness against variations in the input image. We discover that by incorporating FixMatch in the self-training pipeline, the performance improves in the majority of cases. Our performance achieved the best performance with the self-training scheme with consistency on DeepLab-V3 architecture and ResNet-50, reaching 90.69%, 87.37%, and 76.49% on Zheng 1, Zheng 2, and LISC datasets, respectively.