Research And Application Of Medical Cell Images Segmentation Algorithm

Medical image is one of the basis for doctors to understand the condition,diagnose the disease and provide follow-up treatment.Medical image segmentation is a method of dividing useful areas according to certain characteristics of images.It has a positive effect on medical image visualization,lesion area localization,and surgical planning.Among them,the medical image segmentation method based on deep learning has attracted much attention.U-Net is outstanding because it has the advantages of simple network topology and small demand for training set data.However,it will lose segmentation details in the process of segmentation,and the effect is not good for complex images.This thesis focuses on the study of this network,and the specific work is as follows:Firstly,on the basis of retaining the U-Net symmetrical U-shaped structure,a hybrid encoding structure combining ResNet50 and Transformer-Encoder is used.After the encoding,the DAC Block and the RMP Block are added to capture more abstract features and reserve more space information.This thesis proposed a TCU-Net.Experiments on the cell dataset show that compared with U-Net,the Acc,the Recall and the DSC of TCU-Net are increased by 4.16%,9.01%,and 7.55% respectively.Secondly,this thesis analyzes the network structure of U-Net.The two convolution layers of the encoding part are replaced with Residual Dense Block,and the Maximum Pooling structure in the downsampling is replaced with a hybrid downsampling structure,which on the one hand retains the feedforward feature.On the other hand,the features of different scales are extracted,and rich information is selected to give greater weight.Experiments on the cell dataset show that compared with the U-Net,the Acc,the Recall and the DSC are increased by 4.84%,9.63%,and9.51% respectively.It is fully proved that the model proposed in this thesis has a certain effectiveness.For cells that stick together,the watershed algorithm based on distance transformation is used for segmentation,and the effect is remarkable.Finally,this thesis improves the application value of the research content,uses the STM32 development board to design the hardware control system for cell acquisition,and uses the deep learning algorithm to achieve the segmentation and detection of circulating tumor cells.Within the design of the computer code,in addition to the series of functions mentioned above,there are also some operational functions such as doctor’s review and management in the system.These creative functions can not only improve the efficiency of the diagnosis of the disease,but also enhance the practicality of the system.