Human Tissue Cell Image Segmentation optimization Algorithm Based on Improved U-net Network

In medical testing and diagnosis, cells as the basic structure of human body, have received extensive attention in pathological research. The detection and segmentation of cells or nuclei play an important role in describing molecular morphological information. The accuracy of human tissue cells segmentation still needs to be improved, and the network segmentation results have boundary blurring and noise pixels interference. In this paper, an improved U-net network model is proposed. Aiming at the problem of simple stacking of the same convolution operation, a parallel structure for multi-scale image feature extraction is designed. Through the setting of multiple convolution operations and combining different feature fusion methods, a better convolution block structure is obtained. The network segmentation image is optimized through the secondary segmentation of Otsu method and morphological processing, and the final segmentation result is obtained. Finally, the cell contour and centroid are displayed on the original image, and the nuclear center was calculated by image moment. The experiments were carried out on the human organ H&E cell data set. The experimental results show that the Dice coefficient, Jaccard similarity coefficient, recall and accuracy of the improved U-net network model for H&E cell image segmentation reach 0.8260, 0.60, 0.8380 and 0.8270 respectively. Compared with U-net network, the above parameters of the improved U-net model are increased by 1.7%, 3.9%, 1.2% and 2.2% respectively. Compared with CNN network, the accuracy rate is improved by 3.5%, and cell segmentation is more accurate than other literatures.