| With the development of computer science,the combination of artificial intelligence and histopathology has become the a new trend,but the particularity of pathological images makes it face two challenges when using artificial intelligence to assist diagnosis.One is the difference in staining due to factors such as pathological slice preparation process and storage time.These differences can lead to disagreements among histopathologists or AI in the process of diagnosing diseases,which will have a huge impact on diagnostic results.In order to eliminate these differences as much as possible,it is a necessary step to implement a staining normalization method for pathological images.The second is that the amount of data is too small,medical images and other data involve personal privacy,production costs are high,and can not be obtained in large quantities.A large amount of high-quality data in deep learning can improve the model accuracy,while too little data will cause over-fitting problem.In the published data sets,there are not many high-quality pathological images,and even in cooperation with hospitals,there will be lack of data.In this study,three batches of high-quality pathological images of liver cancer were obtained as data sources in cooperation with the cancer Department of triple A hospital in China,and the following studies were made for the above two cases using pathological images of liver cancer.(1)In the study of staining normalization methods for liver cancer pathological images,three typical staining normalization methods,Reinhard,Macenko and Vahadane,were realized according to the characteristics of liver cancer pathological images and the types of stains used.The normalization methods were evaluated according to the image quality analysis index(SSIM).In the improvement process of Macenko’s method,the parameters of three dimensions are searched using the quantum particle swarm method,and the normalized results are evaluated using SSIM.The quantum particle swarm search introduces the average best particle position.At the same time,based on the more randomness of the quantum behavior particle swarm,the optimal normalized parameters can be calculated in a less number of iterations.Compared with Vahadane and Reinhard,the SSIM index was improved to 0.970 after staining normalization using Macenko method with optimal parameters,which effectively solved the problem of staining in the pathological images of liver cancer.(2)In the study of liver cancer pathological image augmentation based on semantic segmentation task,according to the idea of Auto Augment,this thesis proposes an automatic augmentation method of liver cancer pathological image based on semantic segmentation,aiming to calculate the optimal augmentation strategy according to the characteristics of liver cancer pathological image.The automatic augmentation method is divided into three parts: augmenter,sub-network model and controller.The augmenter implements 10 basic image augmentation methods commonly used in deep learning.According to the augmentation strategy,the basic augmentation method is sequentially executed in the corresponding pathological image dataset.The hyperparameters of the augmentation strategy are searched by the random forest based SMBO method in the controller,and then the augmented data are trained using sy BN-based U-net to achieve the maximum iteration.The experimental results showed that the mean values of m IOU and PA of the two batches of liver cancer pathological image datasets were reached 94% compared with the non-augmented liver cancer pathological images using the first batch optimal augmentation strategy,which provided a feasible solution to the problem of small number of liver cancer pathological images. |
PDF Full Text Request