Study On White Blood Cells Segmentation And Classification Of Whole Slide Images

The complete blood count(CBC)is a widely used and significant diagnostic tool in clinical practice,where the white blood cell count is a crucial indicator.White blood cell morphology detection can statistically display the types,numbers,and morphological changes of WBCs,which aids doctors in diagnosing blood-related problems and assists in the diagnosis of diseases.In recent years,fully automated blood cell analyzers based on image technology have been utilized in CBC tests,which have enhanced the efficiency of detection.However,the white blood cell recognition method still has limitations in dealing with special shapes,such as adhesive WBCs and nuclear lobulated cells,resulting in segmentation anomalies and lower accuracy in classifying similar white blood cell types,which restricts its application and makes commercialization difficult.To address these issues,this study conducted comprehensive research and exploration on image preprocessing,white blood cell region segmentation,feature extraction,and classification based on whole-slide images(WSIs)generated by scanning blood smears under an optical microscope.First,a segmentation method based on staining separation and an improved Otsu threshold was proposed to achieve accurate localization and segmentation of white blood cell region,providing high-quality data for subsequent classification and target detection algorithms.First,the image separation was performed using a staining vector to eliminate image fluctuations caused by external factors,retaining white blood cell information that undergoes a chemical reaction with eosin stain and transforming it into a complete region of interest(ROI).Subsequently,the improved Otsu threshold segmentation method,based on the maximum inter-class variance and minimum intra-class variance,was employed to segment WBCs in the image after staining separation.The experimental results demonstrate that this method effectively addresses the under-segmentation of adhesive WBCs and the over-segmentation of nuclear lobulated WBCs,as compared to traditional segmentation algorithms.Second,a feature extraction and classification method based on deep aggregation convolutional neural networks(DACNN)was proposed to achieve accurate classification of white blood cell images.Different convolutional layer features were fused layer by layer from shallow to deep using structures such as iterative deep aggregation(IDA),hierarchical deep aggregation(HDA),and residual functions.As shallow layer features propagate through aggregation nodes at different stages,they are gradually refined to extract more detailed spatial and semantic information,achieving the extraction of fine-grained features and accurate classification of white blood cell images.The experimental results indicate that compared to other CNN models,the DACNN method achieved the highest accuracy rate of 98.73%,effectively addressing the problem of low accuracy rate of some white blood cell types with small differences.Third,a feature extraction and classification method based on a lightweight CNN(LCNN)was proposed.Given that CNN model training parameters are large and prediction time is long,a lightweight CNN model with 10 convolutional layers was designed by reducing feature channels and increasing fusion nodes.This model uses shortcut operations to expedite model optimization,and the fusion of features from the initial,middle,and last convolutional layers is used to achieve a balance between image semantic information and spatial information.The experimental results show that the LCNN method achieves high accuracy(97.28%)while having fewer parameters and a shorter prediction time(3.45 ms)than other CNN models.Fourthly,a real-time white blood cell detection and recognition method based on YOLO improvement was proposed.The method utilizes target detection to achieve detection and classification of WBCs in images potentially containing WBCs as input data.Firstly,the model backbone architecture was built using optimized ELAN,CSPNet,MPConv,etc.and YOLO’s one-stage detection method was employed to simultaneously obtain target position and category through regression,thereby improving the model’s operation speed and feature acquisition ability.Secondly,the model was trained using a dual-head loss structure,three-stage loss function,and adaptive anchor boxes to improve its optimization ability.Finally,model reparameterization was utilized to decouple training and prediction parameters,improving the model’s inference speed.The experimental results show that,compared to the same type of target detection algorithm,under the same experimental conditions,the model achieves the highest average accuracy of 98.14%on multiple test datasets of WBCs,which solves the problem of significant changes in blood smears,and faces the problem of re solving the staining vector for accurate segmentation of WBCs.The model has the smallest number of parameters(38.498×10~6)and obtain the fastest inference speed(19.45 ms).Based on the above experimental results,the following conclusion can be drawn:the aforementioned method can be applied in auxiliary medical devices for real-time classification and counting of WBCs.Compared to current methods,it effectively improves the accuracy and efficiency of leukocyte identification,which is advantageous for the subsequent commercialization of fully automated hematology analyzers.