| Detection and analysis of human cell morphology can reveal cell deformities,proliferation,differentiation,and pathological changes caused by factors such as congenital genetic defects,infections,mutations,etc.,and play a crucial role in evaluating patient health.It is also of great significance in the fields of genetic toxicology and clinical research.Using computer vision technology to segment and study cell samples can efficiently and accurately detect and analyze cell lesions and proliferation,compensating for the subjective impact of detection personnel and the influence of factors such as exposure time and environment on the cell samples in traditional microscopic observation methods,resulting in changes in the characteristics of cells and accuracy of detection results cannot be guaranteed,and time-consuming and labor-intensive problems.This article explores the segmentation of microscopic cell images from both traditional algorithms and the field of deep learning,improving the segmentation accuracy of cell images.The specific research content is as follows:(1)The study background of microscopic cell image segmentation was elaborated in this paper,and the research status of traditional algorithms and deep learning techniques in this field were analyzed from two aspects: domestic and international.Various traditional image segmentation algorithms such as threshold segmentation,watershed segmentation,and active contour model were introduced,as well as classical semantic segmentation model frameworks based on deep learning theory including U-Net,FCN,and Deep Lab v3 plus.The main challenges faced in the segmentation of microscopic cell images were also identified.(2)It was found that traditional algorithms are relatively easy to implement and understand,and are capable of efficiently processing image data.Based on traditional algorithms,a novel cell image segmentation algorithm that uses the color model’s extreme value operation was proposed in this study.Firstly,the S-component image with prominent features in the HSV channel of the image was extracted.Then,the image gradient was corrected using techniques such as morphological reconstruction,H-minima,and image enhancement.The watershed algorithm was then used to segment the image,and the segmentation result was merged into three regions,i.e.,the background,cells,and cell nuclei,based on the grayscale consistency of the original image.Finally,morphological post-processing was performed to remove false noise and flatten the region edges.Experimental results showed that the proposed method achieved high segmentation accuracy for cells and nuclei in human oral mucosal cell images,with accuracies of 0.9867 and 0.9865.Accuracy of cells and white blood nuclei in human blood cell images has reached 0.9469 and 9923,respectively,improving the accuracy of cell image segmentation.,significantly improving the accuracy of cell image segmentation.(3)With the rapid development of deep learning in the field of image segmentation,deep learning models have been applied to microscopic cell image segmentation.These models can automatically learn image features and object classification through end-to-end training,achieving automatic microscopic cell image segmentation.A new encoder-decoder structure model based on deep learning algorithms is proposed for microscopic cell image segmentation.By applying techniques such as dilated convolution and dilated spatial convolutional pyramid pooling to the residual network,the model’s ability to obtain semantic information is improved,and the model’s performance is further enhanced by capturing the multi-scale information of the image.The model also uses cross stage partial networks modules and weighted fusion methods to process features of different levels,obtaining new features with both deep semantic information and shallow positional information.Finally,the segmentation is completed by classifying the feature maps and upsampling them to the original image size.Experimental results demonstrate that the proposed model can achieve accurate segmentation of human oral mucosal cell images and human blood cell images,with an average intersection ratio of 0.8875 and 0.9497,meeting the requirements for precise segmentation of medical microscopic cell images. |
