基于改进U-Net的白细胞提取方法研究

White blood cell extraction method based on improved U-Net

目的:提出一种基于改进U-Net的白细胞提取方法,以解决从血涂片图像中提取白细胞时白细胞粘连和尺寸影响提取精度的问题。方法:在常规U-Net中增加一个形状特征提取层,通过卷积得到细胞的宽度、高度、面积、数量等形状相关信息,并将形状信息引入到损失函数中,从而实现对不同大小的细胞和粘连细胞的有效分割。细胞分割完成后,对分割得到的二值图像进行连通区域提取得到每个白细胞的矩形框区域,从而实现血涂片图像中白细胞的提取。将该方法与U-Net、Faster R-CNN进行对比以验证其提取白细胞的效果。结果:该方法的提取精度(F_(1)值)达到99.20%,较U-Net、Faster R-CNN分别提高了0.80%、2.19%。结论:该方法对粘连细胞及不同大小的白细胞均有很好的提取效果,能够满足白细胞识别分类需求。

Objective To propose a white blood cell extraction method based on improved U-Net to solve the problems of extraction accuracy due to leukocyte adhesion or size in case of white blood cell extraction from blood smear images.Methods A shape feature extraction layer was added to the conventional U-Net to obtain the shape related information such as width,height,area and number of cells by convolution,and the shape information was introduced into the loss function to achieve effective segmentation of cells of different sizes and adherent cells.After cell segmentation connected region extraction was performed on the acquired binary image to obtain the rectangular box region of each leukocyte,thus realizing the extraction of white blood cell in blood smear images.The method proposed was compared with U-Net and Faster R-CNN to verify its effectiveness in white blood cell extraction.Results The extraction accuracy(F_(1) value)of the method proposed reached 99.20%,which was 0.80%and 2.19%higher than U-Net and Faster R-CNN,respectively.Conclusion The method proposed behaves well in extraction of white blood cell with adhesion or with different sizes,and can be used for white blood cell identification and classification.