Abstract：Accurate segmentation of organs or tumors is essential for doctors to diagnose and predict disease.Compared with traditional feature engineering,the classical U-Net model has better performance in biomedical image segmentation.However,in the U-Net model,pooling operation and convolution operation will lose some spatial information,resulting in the decrease of image segmentation.This paper designed the MultiResR2block module to replace the two 3×3 convolution modules used in the U-Net model to extract the features and also designed the PathNet module to connect the encoder and decoder in the MultiResR2U-Net model,so as to reduce the loss of spatial information.At the same time,a new image enhancement strategy is proposed in this paper,so that the end-to-end segmentation model can pay more attention to the parts of the image that are difficult to segment.The segmentation of cell dataset and blood vessel dataset was studied in this paper.The experiment showed that,compared with U-Net and RU-net models,the method in this paper only adopted nearly 2/3 training parameters and obtained better segmentation evaluation performance.In the experiment results of cell segmentation and blood vessel segmentation,the Dice coefficient increased by 0.56%and 1.46%,and the Jaccard coefficient increased by 0.91% and 1.92%.Therefore,compared with U-Net and RU-Net models,our method has better segmentation performance and better generalization performance.
杨晶东,王海灵. 一种有效的全卷积神经网络生物医学图像分割方法[J]. 小型微型计算机系统, 2021, 42(6): 1281-1287.
YANG Jing-dong,WANG Hai-ling. Effective Biomedical Image Segmentation Method Based on Full Convolutional Neural Network. Journal of Chinese Computer Systems, 2021, 42(6): 1281-1287.