WBC Analysis Under Incomplete Supervised Learning

White blood cells(WBC)are an important indicator for diagnosing blood diseases.Currently,doctors mainly count and classify WBC through manual microscopic examination.It is subjective and time-consuming,resulting in deviations in test results or requiring a lot of time.Therefore,it is necessary to study automatic analysis of WBC images based on deep learning.Although data-driven fully-supervised learning methods have achieved significant success,collecting a large-scale WBC dataset is difficult due to the professional annotation,which limits the performance of current supervised algorithms in WBC analysis scenarios.Therefore,this paper aims to study WBC analysis under not fully supervision,exploiting the usage of unlabeled data to improve the performance gain of deep network models.The main research contents are as follows:(1)To address the absence of fast-staining WBC images,this article proposes an unsupervised image style transfer method based on optimal transport.Firstly,this paper explores the optimal transport transfer when cost matrix is color difference,and then propose a WBC image augmentation method by performing optimal transport transfer in subregions.The experimental results show that the proposed method has low computational cost,produces realistic texture for fast-staining white blood cell images without deformation of shape,and achieves similar performance to deep learning-based methods.(2)To address the huge cost of pixel-level WBC images annotation,this article proposes a WBC image semantic segmentation method via entropy minimization.The proposed method utilizes feature gradient regularization,adaptive sharpening,and inter-task class consistency constraints to force the decision boundary to be located in low-density regions.Experimental results demonstrate that the proposed method outperforms comparative algorithms in terms of semantic segmentation performance with minimal spatiotemporal complexity cost on both our constructed blood cell image dataset and publicly available data.(3)To address the huge cost of image-level WBC images annotation,this article proposes a self-supervised WBC classification under heuristic contrastive perspective.This method promotes model to learn the relative relationships between the constructed positive and negative samples,enabling to obtain more regular representations.Experiments on two blood leukocyte image datasets demonstrate that the proposed method can better adapt to various downstream tasks,e.g.,linear classification,domain transfer classification,and fine-tuning classification of WBC images.These methods provide a foundation for developing incomplete supervision-based systems for analyzing blood leukocyte images.