Design And Implementation Of Pathological Slice Analysis Algorithm Based On Deep Learning

With the continuous development of medical technology,cancer is not as incapable as it used to be.If accurate predictions can be made when cancer is in its early stage,the probability of cancer being cured will greatly increase.As the"gold standard"for cancer diagnosis,pathological testing has always attracted the attention of many scientific researchers including doctors.During the diagnosis process of pathologists,fatigue and other factors will inevitably lead to a decrease in work efficiency or even a decrease in accuracy.If the pathologist can assist the pathologist to make a diagnosis through a computer,it will greatly improve the diagnosis efficiency and accuracy of the pathologist.More human lives.In recent years,with the rapid development of scanning technology,especially the emergence of WSI,it has greatly promoted the development of automatic analysis of histopathological images in the computer field.Based on the U-Net model^[48]and the DPN model^[56],this thesis designs a DP-UNet model based on semantic segmentation to achieve cell nucleus segmentation.At the same time,in order to better solve the problem of overlapping cell nucleus segmentation,this thesis adds a distance map module based on the DP-UNet model,that is,while training the semantic segmentation model,train a copy of the shortest distance between the inner area of the nucleus and the outer area.The distance map data is then used to segment the overlapping cell nuclei more accurately through the traditional watershed algorithm.For the nuclear classification task,in order to improve the time efficiency of classification as much as possible on the basis of ensuring the accuracy,this article adds the classification function to the nuclear segmentation model,that is,trains a branch of nuclear classification on the basis of the original model.The pixel-level classification is also used to complete the classification of cell nuclei.According to experiments,the accuracy rate of cell nucleus classification can reach more than 85%.Although the accuracy rate of merging the cell nucleus classification function as a branch into the DP-UNet model is lower than using a convolutional neural network to classify individual nuclei,but the time efficiency has been improved by nearly 5 times,and the accuracy rate basically meets the needs of doctors.In addition to the nucleus segmentation and nucleus classification tasks,this thesis also designs a pathological slice classification algorithm based on the contour information of the nucleus.According to the cell nucleus contour information obtained after the cell nucleus segmentation is completed,cluster analysis is performed to obtain the cell nucleus aggregation area.Subsequently,feature extraction is performed on the nucleus aggregation area,and the extracted features are classified.In the final classifier selection,this thesis tried multiple rounds of control experiments,including SVM,MLP,width learning system,and variant models of the width learning system.In the end,the CEBLS-dense model^[71]performed at best,the highest accuracy rate of its classification can reach 97%.At the same time,in order to better assist pathologists in pathological diagnosis,this thesis designs a pathological slice classification experiment that guarantees zero false negatives.The accuracy of the CEBLS-dense model is still the highest among all classifiers,reaching 90%.