Study On The Segmentation Of Nerve Cell Images Based On Deep Learning And Clustering

The brain is one of the most important organs and controls most of the body's activities.Although medicine develops rapidly in many aspects,there are still no major breakthroughs in the study of the brain's structure and working principles,and it still faces many difficulties.As technology improves,we can easily obtain large-scale,high-quality histopathological images.To systematically study the connections and differences between brain cells and identify pathological cells from them,it is fundamental and critical to accurately segment cells from brain images.The dataset was obtained from the optoelectronic laboratory of Huazhong University of Science and Technology.The data set is a coronal slice of the mouse brain.The purpose of this paper is to accurately segment nerve cells in medical images.However,this kind of pixel-level medical images is very complicated compared to other types of images,such as cell adhesion,blood vessel interference,and foreground background contrast is not obvious.For these reasons,the previous algorithm for segmenting images is not suitable for the data in this paper.This paper aims at the characteristics of microscopic medical images and improves the previous algorithms.First,the pixels in the image are classified by deep learning.Then the results of deep learning training are combined with the original image,and edge detection is performed by clustering.In order to avoid poor fitting effect,in the process of classifying images by deep learning,the image is first enhanced,the sample size is increased,and the noise is increased.Then use a U-net network suitable for segmenting medical images to classify them.In the end,the average accuracy based on the pixel points reached 90.32%,and the average accuracy based on the cells reached 91.60%.Compared with the previous segmentation algorithm with a correct rate of 65% to 85%,this method has higher accuracy.After the distribution of the pixel values in the original image is obtained,thepixel values of the corresponding positions are multiplied by the results obtained by training in the U-net network,and then the edges are detected by clustering.When the number of classes is set to 5,not only can the edge of the cell be accurately identified,but normal nerve cells can also be distinguished from U-net misidentified blood vessel wall cells.