| Pathological diagnosis is an important reference and "gold standard" for cancer diagnosis.Pathological image analysis is the primary means of pathological diagnosis for pathologists.The pathologists make observations under the microscope through the naked eyes and make further judgments based on experience clinically.The results of interpretation are subjective and the clinical misdiagnosis rate is high.In recent years,with the continuous development of deep learning,the automatic analysis technology of pathological images based on deep learning has been promoted increasingly and the objectivity and accuracy of pathological diagnosis have been improved effectively.This paper aims at the clinicopathological diagnosis of two types cancer of gastric cancer and leukemia.Automatic analysis algorithms are proposed correspondingly based on the three pathological data sets.For the task of semantic segmentation of weakly labeled gastric cancer pathological images.Firstly,according to the characteristics of the data set,threshold filtering has been employed by this paper to reduce the effect of false negative data on model training.Then the paper proposes FPA-Net which is based on FPN and ASPP modules.FPA-Net segments cancer area by combines information of multi-scale feature map and various receptive field.Compared with the basic network FPN,FPA-Net can improve the mean Dice coefficient by about 1.41%.For the task of analyzing Whole Slide Image of stomach.First,for the segmentation task,this paper continues to employ the FPA-Net.Second,for further classification of the segmented area,this paper proposes CSResNet,which improves the classification performance of the model by combining attention information of the channel and spatial.Compared with the basic network ResNet,CSResNet can improve the Accuracy,Precision,Recall,F1 Score by about 4.01%,4.27%,1.82%,2.99%respectively.For the task of classification of microscopic image of acute B lymphoblastic leukemia.This paper proposes SDD-Net that can combine the stain quantities that are absorbed by cells and the information in the frequency domain to recognize leukemia cells in microscopic image by utilizing SD-Layer and DCT.Compared with other models in the comparison experiment,SDD-Net has reached the highest of 89.81%and 89.64%respectively in accuracy and weighted F1 Score. |
