Research On Blood Cell Image Detection And Classification Method Based On Deep Learning

Blood routine examination is one of the crucial criteria for evaluating treatment effectiveness,medication,and post-operative recovery.Traditional blood routine examination methods relied on the manual classification of blood cells under a microscope,which is laborintensive and inefficient and prone to errors and omissions impacted by factors such as the personnel’s proficiency in cell morphology and their attention level during the examination.With the continuous development of deep learning technology,an increasing number of neural network models have been utilized for blood cell image processing tasks,and exhibiting high detection accuracy and efficiency.Therefore,this paper employs deep learning methods to detect and classify blood cell images,with the following main tasks:To tackle the task of blood cell detection,this paper proposes an improved algorithm for blood cell detection,named YOLOv5-CBF.First,we introduce a Coordinate Attention(CA)module into the feature extraction part,which significantly enhances the network’s ability to express location features and therefore improves the accuracy of detecting stacked red blood cells.Second,to effectively fuse features of different scales,we incorporate a Bidirectional Feature Pyramid Network(Bi FPN)into the feature fusion structure of YOLOv5,along with an Attention-weighted Feature Fusion(AWFF)module.This improved feature fusion structure enables better bidirectional cross-scale connections and weighted fusion of features,avoiding the problem of overfitting.Finally,we add a small object detection layer on top of the original detection layer to better detect small target platelets and improve the accuracy of platelet detection.Experimental results show that the proposed YOLOv5-CBF algorithm has significant advantages in balancing training time and inference speed,achieving a mean average precision improvement of 2.7% compared to the original network,with a detection speed of 51.9FPS,effectively meeting real-time requirements.This paper presents an improved algorithm for white blood cell classification,called STR50,based on the Swin-Transformer model with a small number of parameters and computational complexity.To capture the features of blood cells,this paper uses the SwinTransformer model’s local self-attention mechanism,while also adopting the concept of a residual network to construct a simple and easy-to-use Res Net50 branch,which,combined with the Swin-Transformer,forms a backbone network structure that fully integrates global and local feature information to avoid the problem of overfitting.Experimental results demonstrate the highly accurate classification of the proposed STR50 algorithm.On the white blood cell dataset of the Kaggle platform,the classification accuracy reaches 99.84%,which is an increase of 1.36% compared to the original network,showcasing great practical value in the field of white blood cell classification.