Mammogram Retrieval Based On Deep Semantic Model

With the progress of society and the development of science and technology,the basic medical facilities are becoming more and more complete,and the clinical medical technology based on image has been widely applied.Effective management and application of large-scale medical image has become an intractable problem in the interdisciplinary field.Medical image retrieval technology,which can retrieve similar cases quickly and effectively,provides effective reference for doctors.It also can be conducive to diagnosis and treatment that help new recruited radiologists and interns grasp the characteristics of medical image and diagnosis experience.Mammography is an important basis for breast cancer screening,but because of the complex shape of lesions such as masses,and the huge difference in human breast density,mammogram analysis and retrieval faces enormous challenges.Based on vocabulary tree that lexicalize image details,this paper regards ROIs in mammography as the research object,deeply studies medical image retrieval based on deep semantic model.The main contributions are as follows:Firstly of all,from the perspective of spatial and semantic information respectively,we analyze the effect of spatial and semantic information on mammogram retrieval,and propose the deep semantic tree retrieval method.The method is based on spatial optimization and semantic optimization respectively,which effectively remove the redundancy of the features by optimizing feature maps from convolutional layer in space or semantic.We also design the adaptive weight for node distribution,then build more streamlined and efficient deep semantic trees to effectively improve the retrieval precision.On this basis,in order to combine the advantages of two kinds of deep semantic trees,we propose a deep semantic tree model based on spatial-semantic dual optimization.For each deep semantic tree,a graph is established between the query image and database images,in which weights are obtained according to the Jaccard index and similarity score and so on,then fuse two graphic models to further optimize the retrieval performance.Finally,in order to further mine the internal relations of deep features,a retrieval method based on semantic micro-forest model is proposed.By constructing different levels of semantic trees,a semantic micro-forest model with deep supervision mechanism is formed,the next layer of semantic trees are constructed with the supervision of tree structure in last layer iteratively,and then the retrieval results of all deep semantic trees will be organic integrated in the retrieval process constantly,which further improves the retrieval precision.The experimental results show that the retrieval methods based on the proposed deep semantic tree and micro-forest model can effectively represent the image semantics,which can mine the intrinsic relations of the deep features of the image significantly.With the proposed methods,we can obtain higher retrieval and classification performance,and they will provide effective technical basis for medical image retrieval research.