Research On Histopathology Image Segmentation Algorithm Based On Weakly Supervised Learning

Distinguishing different tissues in histopathological images is crucial for pathologists to determine the type and degree of lesions.The fully-supervised semantic segmentation algorithm has achieved excellent performance through dense pixel-level labels.However,in the field of histopathology images,it is very difficult to obtain pixel-level labels.To alleviate the dependence of algorithms on pixel-level labels,a popular approach is to use image-level labels to train multi-label classification networks to generate class activation maps and distinguish different tissues.Although class activation maps can reflect the position of tissues in the image,this method has two shortcomings: firstly,as the training deepens,the activated regions will become increasingly concentrated in the most discriminative parts and cannot fully cover the tissues;Secondly,the binary cross-entropy is used to calculate the classification loss of each kind of tissue independently,and different tissue may be activated at the same pixel,resulting in segmentation errors.Given the above shortcomings,this thesis takes the LUAD-Histo Seg dataset as the object to study how to use the image-level label to distinguish different tissues accurately and develops a histopathology image analysis and annotation software.The main research work is as follows:(1)In the process of training multi-label classification networks,this thesis combines channel attention and progressive dropout attention to expand the activation region.Progressive dropout attention dynamically drop the most discriminative regions in the feature map based on the training epochs.The deeper the training,the larger the discarded regions and the network needs to explore more regions to maintain classification prediction.At the same time,we add channel attention before channel attention,obtain the averagepooled feature and max-pooled feature through global pooling operation on the feature map,input two one-dimensional features into the multi-layer perceptron to generate channel attention map to capture the importance of different channels,and finally multiply the channel attention map and the input feature map channel by channel to guide the network to pay attention to important features.(2)This thesis designs and implements a class reactivation mapping method.After the multi-label classification network converges,we generate a single-label feature and train an additional fully connection layer.Through the class mutual exclusion feature of softmax cross-entropy loss function,the network punishes the wrongly activated regions and improves the network segmentation performance.(3)Using PyQt5,OpenCV,Pytorch and other frameworks,we have developed a histopathology image analysis and annotation software that integrates deep learning models.In addition to basic image processing functions,the software also integrates deep learning models.Annotators can convert the segmentation results generated by the model into editable polygon primitives,and then complete data annotation by fine-tuning the polygon primitives to reduce data annotation costs.The segmentation method proposed in this thesis has been tested in the LUADHisto Seg dataset,and reached 76.53% MIo U,which outperformed the existing methods.At the same time,the effectiveness of combining channel attention with progressive dropout attention and class reactivation mapping was demonstrated through ablation experiments.The designed algorithms and software help promote the automation of medical image analysis technology in pathological research and application.