Research On Classification Of Static Histopathology Images

Pathological diagnosis is recognized as the "final diagnosis" of the disease because of its accuracy.The traditional pathological diagnosis,which is done independently by pathologists,takes a long time and is highly subjective.With the development of electronic imaging technology and artificial intelligence technology,the combination of artificial intelligence technology represented by deep learning and traditional pathological diagnosis has been widely studied.The use of deep learning technology to assist pathological diagnosis can not only improve the accuracy and efficiency of pathologist diagnosis,but also greatly improve the uneven distribution of medical and health resources in China.However,most of the relevant research is based on full-load slide digital scanning technology to obtain WSI images,application of scanning equipment and data storage have a high demand,the accumulation of a large number of clinical pathological images(referred to as clinical pathological images)are not fully utilized.In clinical practice,pathologists use microscopes to observe slices and intercept a few images as a basis for diagnosis,and apply deep learning techniques to process clinical pathological images are challenged as follows: First,the clinical pathological images preserved as diagnostic basis have mixed,unknown scales,and complex and variable staining conditions,making it difficult for depth models to fully learn the characteristics of tumor cells in the entire scale space and staining space;Risk of overfitting and significant data loss,and third,pathological images and diagnostic information are usually stored on a patient-by-patient basis,resulting in problems where a single pathological image cannot correspond to specific diagnostic information.In view of these problems,this paper first studies the data enhancement techniques for clinical pathological images,then further studies the classification algorithm for class imbalance-oriented clinical pathological images,and finally further studies the classification algorithm for clinical pathological images that do not correspond to labels,and the main research progress is as follows:To solve the problem that mixed and unknown scale and staining of clinical pathological images is the key and foundation of applying deep learning techniques to process clinical pathological images.Most of the existing multiscale processing methods are based on known or controllable image scales and cannot handle the mixed and unknown scales and staining conditions in the original clinical pathological images.To this end,the Online Adaptive Data Augmentation(OADA)is proposed.In order to fully learn the characteristics of the pathological tissue in the entire scale space and staining space,OADA first divides the base image candidate set from the training set,then adaptively selects the base image from the candidate set according to the current training state at each training stage,and determines the personalized enhancement quantity of each selected base image in conjunction with the current and historical training state,which is used to update the training set with data enhancement.The experimental results show that OADA-enhanced depth models have an average classification accuracy of 4.78 percent on scale and stained mixed pathological images compared to the existing data enhancement methods.In the task of medical imaging classification,due to the difference between the number of negative and positive samples and the difference in the number of disease subtypes in the positive sample,the imbalance of different degrees of category is an unavoidable problem.There is a large imbalance in the number of pathological images of scale cancer and adenocarcinoma in the data set of lung cancer subtypes,and there are also problems of mixed scale and staining in these images,and the accuracy of the existing problem-solving strategy classification of unbalanced classification is not ideal.To this end,this paper presents the Model Ensemble and Curriculum Learning-based Classification for unbalanced clinical pathology images(MECLC).MECLC first divides the unbalanced classification data set into k more balanced base subsets,then trains the depth model separately on the base sub-data set,looks for the easy-to-recognize images and difficult-to-recognize images in the data set,and trains the depth model from easy to difficult based on the idea of course learning,and finally integrates the trainingacquired depth model.The experimental results show that the classification accuracy obtained by MECLC is improved by an average of 9.13 percent compared with the baseline algorithm and the existing problem-solving strategy of unbalanced classification.In a hospital’s clinical database,data is usually stored on a patient-by-patient basis.Multiple pathological images of patients support different diagnostic information,but how the pathological images and diagnoses correspond is unknown.Exploring the correspondence between pathological images and diagnostic information,i.e.finding pathological images that support the diagnosis of lung cancer subtypes,can help expand the database of pathological images and promote the application of information technology represented by deep learning in clinical medical research,but there are few relevant studies at present.In this paper,the Multiple Instance Learning and Attentionbased Classification for Pathological Images(MILAC)is proposed.MILAC refers to the multi-sample learning model,organizes the pathological images into packages in patient units,and takes the patient’s lung cancer subtype diagnosis as the label of the package,designs a depth model with attention mechanism,uses the VGG16 network as a feature extractor,initializes and updates the weights of the pathological images in the package according to the network attention and feature distance,and realizes the corresponding patient level labels and pathological image labels according to the weights.The experimental results showed that MILAC’s classification accuracy of pathological images in patient packages improved by an average of 114.73 percent compared with baseline algorithms and existing multi-sample learning algorithms.