| With the continuous development of the social economy,the material conditions are becoming more and more superior,but it also leads to many unhealthy living habits,resulting in an increasing number of cancer cases in recent years,of which breast cancer accounts for the highest proportion of new cancer cases among women.It is one of the most effective ways to prevent breast cancer by using mammography to detect early symptoms of breast cancer.Using computer-aided diagnostic tools to help doctors screen for breast cancer can not only alleviate the pressure of doctors,but also reduce the requirement of doctors’ ability to read films.Retrieval of breast mass mainly uses the morphological and structural features of mammography,and performs similarity search in the image database to find cases similar to the query images,so as to provide a reference for the diagnosis of breast cancer.Inspired by the deep hash-based image retrieval method,combined with the idea of joint feature learning,the paper makes full use of the similarity and difference between image features,and proposes a deep hash mammography retrieval method based on joint feature learning.The main work is as follows:Firstly,this paper proposes a deep hash retrieval method based on joint feature learning.In order to solve the problem of similarity between normal tissue image features extracted by convolutional neural network and some early lesion image features,a method of simultaneous learning in feature extraction stage is proposed,which learns similarity between similar features and difference between different features at the same time.The method can increase the feature distance between normal tissue and pathological tissue,and solves the problem that similar features can not be clustered better and the recognition degree of different features is not high in the process of extracting hash features by convolutional neural network.The hash feature obtained by the algorithm keeps the original similar relationship with the image to be retrieved,and effectively improves the retrieval precision.Secondly,the paper proposes a similarity-difference-based benign and malignant retrieval method for breast masses.The method introduces the vector norms to represent feature distances to deal with the problem that the hash characteristics of benign and malignant masses are close to each other in Hamming space.In the feature learning stage,the vector similarity difference is used to increase the distance between dissimilar features to avoid the vector inner product cannot represent the Hamming distance between vectors in an all-round way,and achieves the purpose of better distinguishing the nature of lesions.The algorithm has performed well in the benign and malignant retrieval tasks of breast masses,and effectively improves the retrieval accuracy.Finally,this paper builds an assistant image retrieval platform for patient diagnosis and treatment process.Different from the most internet medical software platforms usually only have the functions of registration,disease supervision,doctor-patient communication and so on,the platform provides patient-oriented image retrieval function for diseases that need to be diagnosed by filming.By retrieving similar cases,the platform can timely understand the condition of patients,and solve the problem that patients cannot accurately describe the real condition in language,which result in the failure to obtain professional diagnosis and treatment opinions.Through the construction of the assistant image retrieval platform,the effective combination of algorithm theory and practical application is realized.The experimental results show that the proposed deep hash mammography image retrieval method based on joint feature learning can effectively improve the accuracy of breast mass retrieval.It can efficiently conduct retrieval of normal tissue and pathological tissue,also benign and cancer masses,which can provide a good theoretical and technical foundation for the retrieval task in clinical medical scenes. |
PDF Full Text Request