Analysis Of Lesions In WSI Of Breast Cancer Based On Deeplearning

The number of women who die from breast cancer every year is second only to lung cancer.Breast cancer is the most common cancer.Early detection of breast cancer and treatment measures can significantly reduce the mortality of breast cancer.Patch based analysis of lesions in WSI of breast cancer is widely used in breast cancer detection.This thesis mainly focus on the three aspects of the imbalance of training data,the spatial information in WSI and the acceleration of inference.There are a lot of negative samples and easy samples in the training dataset.The imbalance of positive and negative samples results in the model’s bias which reduces the accuracy of the model.A large number of simple sam-ple gradients can easily overwhelm the difficult samples which have higher learning value,resulting in slow learning rate and reduction of model perfor-mance.In this thesis,we will discuss data sampling,loss function based on category weight,Focal Loss and GHM.These methods were compared base on the dataset of breast cancer.Aiming at the problem of spatial information loss,we propose a spatial information fusion framework called CNN-LR.According to the structure of deep learning model,CNN-LR divides the classification process into feature extractor and classifier.As a feature extractor,CNN can extract detailed in-formation of the image and LR receives the contextual information of multiple adjacent images as input.CNN-LR reestimate the probability of patch base on the larger receptive field.In order to avoid the redundant calculation of the overlapped image area in the inference stage,we propose a detection method based on the full convolution structure.In this method,full connection layer is transformed into an equivalent convolution structure,which avoids the limitation on the resolution of the input image.By increasing the input image size in the inference stage,prediction of multiple image regions can be achieved in one inference.Feature extraction results of overlapping regions are shared,which avoids redundant calculation and greatly accelerates the speed of model inference.