Semi-Supervised Semantic Segmentation using Redesigned Self-Training for White Blood Cells

Artificial Intelligence (AI) in the medical field, especially in the diagnosis of cancer relating to white blood cells, is held back by two primary obstacles: the scarcity of extensively labeled datasets for white blood cell (WBC) segmentation and outdated segmentation techniques. These challenges stall the development of more accurate and modern techniques to diagnose cancer relating to white blood cells. To overcome the first challenge, there is a need to devise a semi-supervised learning framework to efficiently make use of the huge dataset that is not annotated. Our study tackles this by introducing an innovative self-training approach that integrates FixMatch. Self-training is a semi-supervised framework that leverages the model trained on labeled data to generate pseudo-labels for the unlabeled data and then re-train on both of them. FixMatch is used to regularize the model against weak and strong perturbations in the input image. We discover that by incorporating FixMatch in the self-training pipeline, the performance improves in many cases. When tested, our method achieved superior results on the UNet and DeepLab-V3 networks with ResNet-50 architecture, attaining accuracies of 90.87%, 89.05%, and 75.13% on the Zheng 1, Zheng 2, and LISC datasets, respectively.