Detection And Classification Of Nuclear Image Based On Convolutional Neural Network

The diagnosis of pathological tissues is the basic standard for diseases such as tumors and cancer,and histopathology is a typical tool for diagnosing cancer.Correct detection and judgment of the nuclear class play a central role in treatment decision-making,mainly by pathologists examining the image of the tissue specimen and defining information reports about the appearance,cell composition,disease or normal state of the tissue.Using techniques such as image processing and machine learning,pathologists can perform automated nuclear detection and classification of images to make quantitative and qualitative assessments of diseases,not only for pathologists to reduce the heavy workload,but also to confirm the diagnosis and the result is more objective.However,since most of the nucleus of the pathological image is clustered and its histological structure exhibits complex and irregular morphology,it brings some challenges to the first detection and classification of the nucleus.After summarizing and analyzing the traditional methods and deep learning methods,this paper designs a convolutional neural network based algorithm for the detection and classification tasks based on the characteristics of the nucleus in colorectal adenocarcinoma tissue images.The main work of this paper includes:1)Aiming at the problem that the number of nuclei in the pathological images is uneven and the shape is small,a nuclear position detection network with hierarchical feature fusion is proposed to explore the mapping relationship between the nuclei and their position maps in the pathological images.Compared with the existing detection algorithm,the sliding frame is avoided and the parameters are reduced;The pre-training model in the down-sampling process is designed for the sample to be unbalanced and less,and the cascade module of multi-level complementary information in the up-sampling is responsible for refining the positional response of the restored nuclei,and the predicted position probability map is partially post-processed.After that,the predicted coordinates of the obtained nuclei will be calculated.2)Aiming at the problem of nuclear center image offset and less supplementary information,a classification network of regional cutting module and context feature supplemental network is proposed,which provides the complementarity of local and global features for the classification process of nuclear.Compared with the existing classification algorithm,the multi-channel feature and multi-directional feature fusion are performed,which avoids the influence of single feature extraction on classification accuracy.The region cutting module has the advantages of multi-directional feature extraction and strong flexibility,while the context supplemental network provides the semantic information of the surrounding background for the limited nuclear image,and the two complement each other,and the final fusion obtains the classification label of the nuclear.3)According to the above model,this paper carried out the verification experiments of single detection,single classification and combined detection classification framework on the histopathological dataset of colorectal adenocarcinoma.The experimental results of each group are higher than other comparison methods.The position detection network improves the accuracy of the location of the nucleus in the complex background.The proposed classification network improves the classification accuracy of the nucleus,and the experimental results of the combined detection and classification network are greatly improved.