Research On Mammography Image Retrieval Method Based On Deep Learning

In recent years,breast cancer has become one of the high-risk cancers that seriously threaten women's health.Early detection and early diagnosis are currently the best methods for treating breast cancer,and molybdenum target X-ray imaging is currently the most suitable method for general screening.He research work in this paper is mainly based on deep learning methods for the diagnosis of early breast cancer,and explores the application of deep learning methods in breast cancer mass retrieval.In the past,most breast cancer mass retrievals used machine learning methods based on manual feature extraction,which had significant limitations in the semantic and visual similarity of retrieval results.In recent years,deep learning has achieved breakthrough development in the field of computer vision and achieved excellent application results.In this study,we carried out a deep learning-based mammography target retrieval method study.The specific research content includes the following two aspects:(1)Research on mammography retrieval based on deep learning feature expressionUse three deep learning frameworks to pre-train benign and malignant binary classification of breast cancer region of interest(ROI),select the feature expression with the best classification effect as the feature vector for retrieval search,and finally calculate the ROI to be diagnosed The Euclidean distance from the ROI feature vector in the confirmed image database is used for retrieval.From the research results,the Resnet model can better extract the feature expression of benign and malignant breast cancer.By calculating the distance of the feature expression,the average retrieval accuracy and average retrieval accuracy(evaluation index of perceived similarity)are0.90 and 0.62.(2)Research on mammography retrieval based on deep learning similarityAccording to the "gold standard",the perceptual similarity between ROIs is scored,and then the perceptual similarity between ROIs is learned through a two-channel convolutional neural network,and then retrieved based on the perceptual similarity scores output by the model.The average classification accuracy and average retrieval accuracy of benign and malignant retrieval of mass ROI based on deep learning perceptual similarity method reached 0.70 and 0.76,respectively,and the average classification accuracy was improved to 0.74 after weighting,and twooptimizations were carried out subsequently The average classification accuracy was improved to 0.90 and 0.73,respectively.The research results show that the deep learning-based retrieval method can provide a similar ROI in both semantic and perception.Through further optimization and organic integration with clinical practice,it can help improve the doctor's diagnostic accuracy and have potential clinical Application prospects.