Research On Method Of Breast X-ray Image Description And Pathological Image Classification Based On CNN

Breast cancer is one of the most common malignant tumors in women.Early detection and early treatment are important to reduce the mortality rate of patients.However,at present,there is insufficient supply of high-quality domestic medical resources,high medical costs,long professional doctor training cycles,and high work intensity,easily lead to a higher rate of misdiagnosis.In view of such medical conditions,this paper uses deep learning technology as the theoretical basis,and uses the convolutional neural network(CNN)to study the text classification method of breast X-ray image description and the image classification method of pathological images to assist doctors in the diagnosis of breast cancer process to improve the doctor’s work efficiency and reduce the rate of misdiagnosis due to fatigue and other factors.First,the initial screening of the patient is completed by X-ray examination of the breast to determine whether the patient has the possibility of malignant lesions;then,further medical examinations are performed on patients with the possibility of malignant lesions,such as biopsy to obtain pathological images to complete the benign and malignant judgments and tumor subtypes to achieve the final diagnosis.First,aiming at the mammography description classification problem,the scheme is designed according to the characteristics of the text classification task.According to the task requirements,collect the mammography report,establish a text case set and complete the analysis and arrangement of the dataset.Extract image description text data based on OCR technology,establish data cleaning rules to remove noise in text case data;use Chinese word segmentation technology to complete data feature extraction,establish a custom dictionary,and improve the accuracy of word segmentation results;use distributed representation method to represent the feature of words and text,and through the above preprocessing process,medical texts expressed in natural language are converted into high-quality structured data,which is convenient for computer to read and calculate.Secondly,according to the characteristics of the text classification problem,complete the design of the input layer,convolutional layer,pooling layer,fully connected layer,and output layer,and complete the construction of the image description text classification model based on the one-dimensional convolutional neural network.The training set completes the training process of the model,establishes an accuracy evaluation index,and uses experimental verification to complete the selection of the model structure and the adjustment process of the hyperparameters,which effectively improves the classification accuracy of the model.Then,according to the breast pathology image classification problem,Brea KHis,a public data set of breast pathology images,is used to sequentially perform image preprocessing such as normalization,size scaling,and image cropping.According to the characteristics of image classification system,an image classification model is designed based on two-dimensional convolutional neural network.Based on the idea of transfer learning,two different migration methods are designed to complete the pre-training process of the pathological image classification model.Experimental verification is performed for the two classification tasks of judging benign and malignant tumors.Through experimental comparison,we can know that the method of fine-tuning the network weight can obtain higher accuracy of classification,meanwhile effectively save the calculation cost and shorten the time of model training.Finally,for the problems of too many parameters in the fully connected layer of the original model structure,low calculation efficiency,and the small size of the dataset,which easily leads to overfitting,the model structure and data augmentation methods are designed separately.Through comparative experiments,the effectiveness of the improved structure and data augmentation method is verified.The number of parameters included in the improved model structure is greatly reduced,and the computing efficiency is significantly improved.At the same time,the size of the dataset is increased,and the adverse effect of the overfitting phenomenon is effectively suppressed,which makes the model have strong generalization ability.In the end,after experimental tests,the above two methods can significantly improve the classification accuracy of the model in both 2 and 8 classification tasks.At the same time,by comparing recent research results at home and abroad,we can know that the pathological image classification model designed in this paper has obtained Up the leading level.