Research On Classification Method Of Pathological Images Based On Deep Learning

As the "gold standard" of clinical cancer diagnosis,pathological diagnosis is of great significance to determine the treatment plan and prognosis of patients.Traditional pathological diagnosis uses manual slides reading,which has some problems,such as low accuracy,heavy workload and so on.In order to solve these problems,computer-aided diagnosis technology came into being.It can help doctors quickly find abnormal tissue areas from massive data and improve doctors’ diagnosis efficiency and accuracy.Pathological image classification is a key technology in the field of pathology aided diagnosis.The classification performance of the existing models can not meet the requirements of practical diagnosis.This is due to the high resolution and high labeling cost of pathological images and the tumor tissue itself has intra class variability,its morphological and texture features are very complex.In recent years,deep learning technology has been widely used in the field of computer vision and has shown great potential in the field of pathology aided diagnosis.Therefore,This paper uses deep learning method to solve the problems of pathological image classification.The main research contents are as follows:1.For the classification task of pathological WSI,A weakly supervised pathological WSI classification method named MoCo-CLAM based on contrastive learning is proposed,which solves the problem that CLAM does not make full use of the data in the domain of pathological image.MoCo-CLAM first extracts patches from the tissue region in the WSI,then uses the encoder to extract the features of patches,and finally uses CLAM to aggregate all features and classify WSI.The encoder is trained by MoCo v2 and intra domain data in an unsupervised manner.When we only have WSI labels,MoCo-CLAM has better classification performance than the performance of other weakly supervised methods and some supervised methods on two kinds of WSI data sets.2.For the classification task of small-size pathological images,A classification model named Spatial＆Stain Attention Network（SSANet）of pathological images based on color deconvolution is proposed.Combined with stain separation technology,We propose a Spatial＆Stain Attention mechanism so that the network can focus on the stained nuclei and cytoplasm in the pathological image.SSANet firstly uses the ACD to generate the stain separation image,then uses the stain separation image to activate the SSA mechanism during training,and finally uses a linear classifier for classification.Experiments show that the classification performance of SSANet in CRC-HE dataset and Break His dataset is better than the performance of other models.3.For the lesion region detection task of WSI,A detection method of lesion region in WSI based on SSANet is proposed.In this method,the SSANet is transferred to the lesion region detection task,and the pixel segmentation problem is transformed into an patch classification problem.The lesion region detection can be implemented only by using some patch-level labels.This method firstly trained the SSANet,then used the trained SSANet to predict the patches of the test set,and finally used the predicted results of patches to generate the heatmap of high-risk lesions region.Experiments show that our method can effectively detect suspicious lesion areas in pathological WSI.To sum up,starting from the diversity and complexity of pathological images,this paper puts forward specific solutions to the problems existing in the above tasks.The experimental results show that the proposed method performs well in the corresponding tasks.