Research On Medical Cell Image Segmentation Based On Convolutional Neural Network

With the number of cancer patients increasing year by year,it is of great significance to use computer aided diagnosis system to help doctors determine whether cells become cancerous.Medical cell image segmentation is an important part of computer aided diagnosis system.There are many problems in cell image,such as low contrast between nucleus and image background,large difference in nucleus morphology,overlapping of nuclei and complex image background.Those make it difficult to accurately segment nucleus.Therefore,a thesis proposes the D₂U-Net with multi-scale feature extraction module and dual decoder.It is combined with watershed algorithm to achieve precise nucleus segmentation.To solve the problem of nucleus segmentation,a D₂U-Net with dual decoder is proposed.The network takes U-Net as the basic network,and it introduces multi-scale feature extraction module and nucleus centers prediction branch.The multi-scale feature extraction module extracts image features of different scales and deeper levels.The prediction branch of nucleus centers plays an auxiliary role in the segmentation of nucleus by the main branch,making the main branch more accurate in the segmentation of nucleus.Experimental results on the cervical cell dataset and the Mo Nu Seg dataset show that D₂U-Net is effective,and D₂U-Net has better segmentation performance than U-Net and its variants.To solve the problem that nucleus and nucleus centers occupy a small area in the image,a hybrid loss function is proposed.The loss function consists of Dice loss function and BCE loss function.By adjusting the proportion of the two in the mixing loss function,the influence of unbalanced proportion of positive and negative data in the dataset sample can be eliminated.Experimental results show that the mixed loss function significantly improves the scores of Dice,Precision and Recall compared with Dice loss function and BCE loss function.To solve the difficult problem to segment overlapping nuclei,a post-processing segmentation method based on watershed algorithm is proposed.The nucleus centers predicted by D₂U-Net and the nucleus centers extracted by threshold segmentation were used as markers of watershed algorithm respectively.Then two watershed transformations were carried out.The results of two overlapping nucleus segmentation were fused with the nucleus segmentation results of D₂U-Net.Experimental results show that the first and second watershed transform and the method of extracting the nucleus centers can effectively segment overlapping nuclei.In addition,the segmentation method combined with D₂U-Net and post-processing has higher scores in AJI,Dice and PQ than most of the current methods,and the segmentation effect is good.