The Research On The Algorithms Of The Pathological Image Segmentation Based On Gabor And Cascade Squeeze Bi-Attention Strategy

Histopathology image diagnosis is the " gold standard " for tumor diagnosis,and accurate segmentation of the lesion area from histopathology images is a crucial step to assist pathologists in diagnosis.However,the limited number of pathologists read slides,the diversity of tissue lesions,and the differences in the diagnosis of different pathologists have brought difficulties to tumor diagnosis.Computer-aided diagnosis system can reduce the workload of pathologists and help pathologists diagnose faster and more accurately.Therefore,designing an accurate automated segmentation algorithm is of great significance to assist pathologists in diagnosis.The typical image segmentation method forms a dense image representation,ignoring its texture and multi-scale attention information.Based on this,this paper proposes a GCSBA-Net(Gabor-based and Cascade Squeeze Bi-Attention Network)algorithm for extracting multi-scale,multi-directional texture information and attention information from histopathology images.Firstly,the deep learning method has weak interpretability,poor robustness,and large limitations.This paper establishes a Gabor-based auto encoder module to make up for the deficiencies of deep learning methods.Specifically,this paper uses the ideality of Gabor wavelet for texture feature extraction,and modulates the Gabor filter to the CNN convolution kernel to extract features similar to the Gabor filter.Experiments show that the module not only extracts texture information of different scales and directions,but also has a certain interpretability.Secondly,in the process of CNN extraction,it is easy to lack the local information of the image and ignore the detailed features in the image,which causes the semantic information to be blurred.In response to this problem,this paper designs a Cascade Squeeze Bi-Attention model(CSBA),which combines cascaded features and attention mechanisms and applies them to convolutional neural networks.The model contains the Atrous Cascade Spatial Pyramid Module(ACSP)to extract features of four different scales,and each scale branch contains a Squeeze Position Attention module(SPA)and a Channel Attention module module(SCA)to maintain the multi-level aggregation on the spatial pyramid with different dilations.Experiments show that the cascaded squeeze bi-attention module can mine hidden details to improve the detection accuracy of the network.Finally,for the problem of imbalance of data distribution and boundary blur in image segmentation,this paper proposes a boundary-aware hybrid loss function to better respond to the boundaries of the images.The boundary-aware hybrid loss function measures the difference between the input image and the annotation.Through Binary Cross Entropy(BCE),Structural Similarity(SSIM)and Fβ-focal loss function,the fusion of pixel,detail similarity and sample balance is achieved.Combined with the boundary-aware hybrid loss function,the proposed GCSBA-Net can effectively segment the salient target area and accurately predict the clear boundary structure.The method in this paper is trained,tested and evaluated on the GlaS dataset of the MICCAI 2015 Challenge and CRAG dataset,and compares the results of existing benchmark models on evaluation indicators such as F1 score,Dice coefficient and Hausdorff distance.The performance of the GCSBA-Net proposed in this paper is better than FCN-8,U-Net,SegResNet and Deeplabv3+Networks,and has achieved state-of-the-art performance.In addition,for the problem of gland segmentation in pathological images of rectal cancer,this model can accurately extract smooth gland edges and the closest approach to the pathologist’s annotation.