Unsupervised Refocusing Method Of Pathological Images Based On Attention Mechanism And Domain Normalization

As the main object and carrier of pathological analysis and diagnosis,digital pathology images have very important research and application value in the field of biomedicine.Due to the uneven distribution of cells on the smear and the limitations of the autofocus algorithm,it is almost impossible to maintain all cells within the depth of field of the objective lens during imaging,which causes the defocused cells on the digital pathology images.Defocus can cause blurred cell structure,which greatly affects pathologists' reading and the application of automated algorithms based on digital pathology image.The existing methods for solving defocused image blur are mainly multi-focus image fusion and single image deblurring.Multi-focus image fusion algorithm needs to obtain images under multiple focal points in advance,which are high requirements for hardware and time-consuming.Current single image deblurring methods mostly use supervised training methods,which are difficult to generalize in the unlabeled data domain;even if there are a few unsupervised methods,they generally target global motion blur,which is difficult to apply to sparse pathology images with out-of-focus blur.In response to the above challenges,this thesis proposes an unsupervised refocusing method of pathology images based on attention mechanism and domain normalization.(1)Aiming at the problem that refocusing is not sensitive to sparse defocused areas,this thesis proposes a refocusing method for defocused images based on the attention mechanism.This method uses the generative adversarial network(GAN)model to reduce the distribution difference between the generated refocused image and the real in-focus label image in the confrontation between the generator and the discriminator to continuously improve the reconstruction effect of the model.For the locality of the defocus blur spatial distribution and the discreteness of the defocus degree,a defocus mask segmentation network is designed to enhance the attention to the local defocus area in this thesis;and further designed a novel maximum image patch GAN loss,which strengthens the refocusing model's ability to restore local texture details.In the network structure,the ordinary U-Net is improved and the dense connection convolutional structure is introduced to strengthen the characteristic expression ability of the network structure.(2)Aiming at the problem of model generalization,this thesis proposes an unsupervised refocusing method based on domain normalization.This method can complete the generalization training of the unlabeled data domain without label supervision,and obtains pretty good results.The main structure is composed of domain normalization and refocusing.The domain normalization part normalizes images of different domain styles to the same domain space to reduce domain differences,and introduces red and blue mask loss and gray mask loss to prevent red and blue cross colors and image distortion.In the refocusing phase,this thesis will conduct pre-training on the labeled data domain,then the hybrid training is performed on the normalized multi-style data field to realize the refocusing on the unlabeled data field.Multi-scale nuclear and cytoplasmic feature branches are designed on the network structure,which improves the restoration and reconstruction of nuclear and cytoplasmic details.(3)Applying the above refocusing method to pathology images,this thesis demonstrates the enhancement of the image quality of entire digital pathology image and the improvement of the effect of cell nucleus segmentation.The experimental results prove that the method in this thesis is significantly better than the existing methods in terms of structured similarity,peak signal-to-noise ratio and standard deviation and other image quality evaluation indicators.Compared with the defocused input image,the average peak signalto-noise ratio has increased by 10.21%.For nuclear segmentation tasks,this thesis evaluates the index of the nuclear segmentation on the test data before and after the refocusing.The average intersection over union increases by 12.73%,which effectively improves the effect of the nuclear segmentation task.As mentioned above,in order to solve the problem of sparse defocus in pathology images and the difficulty of generalization of models,this thesis proposes an unsupervised refocusing method based on attention mechanism and domain normalization,which effectively solves the problem of defocus in pathology images.Further application in practical tasks proves that the refocusing method has important practical application value.