Research On Multi Staining And Multi-scale WBC Instance Segmentation Algorithm

Leukocyte is an essential part in the human immune system.The type and number of leukocytes provide important information about human health.In recent years,the automatic detection of leukocytes have been one of the representative topics in medical image processing.Some existing networks for medical image segmentation have solved some problems.However,there is no more suitable method for multi-scale and multi staining leukocyte microscopic images.To solve the above problems,this paper has done the following work:In this paper,a new leukocyte data set is established,which is composed of two parts.One part consists of 302 leukocyte images provided by Tianjin Medical University;The other part includes 242 leukocyte images from the public dataset LISC(Leukcyte Images for Segmentation and Classification).There are totally five categories of leukocytes: neutrophils,lymphocytes,monocytes,eosinophils and basophils.For this dataset,this paper adopts the basic instance segmentation framework,which is composed of feature extraction module,feature pyramid module,region generation network and a detection network.For a variety of staining situations,this paper adopts a feature extraction network that can strengthen feature expression.Through feature extraction from multiple subspace and strengthening the weight of the specific channels,it can alleviate the impact of different staining on the essential features of cells and extract more diverse leukocyte features,so as to strengthen feature expression.In the process of feature fusion,the salient features of each scale are highlighted by strengthening the features of different scale feature layers,so that the network can recognize multi-scale cell targets more accurately.Compared with other target detection and instance segmentation methods,the network structure in this paper achieves better results in accuracy and other aspects.According to the experimental results of the above model,this paper makes further optimization.Considering that the poor detection results are resulted form the small-scale similar cells,the improvement is mainly carried out from the following aspects: firstly,the small-scale features are supplemented,the anchor matching strategy is adjusted,and additional suppression operations are carried out on the redundant negative samples.Then,due to the imbalance of classification results,specific training strategies are formulated for difficult classification tasks.The experimental results show the effectiveness of the improved model and the necessity of each module.