| When using traditional deep learning for pathology image classification,researchers often consume a lot of time and labor costs due to the need to label a large number of pathology images.In order to reduce the time and labor cost of labeling images,this thesis studies existing solutions.When classifying histopathological images,people will crop the whole slide image(WSI)into many small patches because of the huge WSI.While cropping the patches,the doctor labels each patch.This allows these patches to be fed into the neural network for training.The job of labeling patches is time-consuming and labor-intensive for doctors.The researchers thought of using the Multiple Instance Learning(MIL)attention mechanism to solve this problem.However,the MIL attention mechanism is a hard attention mechanism and cannot evaluate the weight inside each patch,thereby reducing the final classification accuracy.This thesis proposes the use of two methods to address the lack of soft attention mechanisms for MIL attention mechanisms,and demonstrates the superiority of both methods by classifying them on the colon cancer dataset.The first method proposed in this thesis is a soft,sequential,spatial,top-down attention mechanism(S3TA).This attention mechanism is inspired by the primate visual system.Rather than viewing an image as a static scene,humans explore the image in a series of glances,gathering and integrating information in the process.This is why humans are always better at classifying pictures than usual neural networks.S3 TA is a recurrent neural network(RNN)that simulates the attentional bottleneck and function of sequential,top-down,recurrent control of the visual cortex.Adding S3 TA to the MIL attention mechanism is beneficial to make up for the insufficiency of the hard attention mechanism,and can also obtain the advantages of human image classification.The second method proposed in this thesis is to improve the MIL attention mechanism from the channel domain perspective.This thesis uses a squeeze and excitation(SE)attention mechanism module to improve the MIL attention mechanism.The SE module represents the interdependence of each channel in the convolutional layer,so that the network can recalibrate the weights of each channel domain,that is to say,the SE module can emphasize important information and suppress invalid features in the channel domain.Because the depth of all network models in this thesis is shallow and the original SE module performs poorly in shallow layers,this thesis improves the SE module.The improved SE module can not only insert the network model at any depth,but also retains the advantages of simplicity,flexibility and efficiency.The experimental results show that the two new models have higher classification accuracy than the original MIL attention mechanism model,and the combined model performs better.This thesis provides a more effective solution for reducing the cost of WSI labeling. |
