Research On Medical Image Retrieval Based On SAE Feature Extraction

With the development of science and technology, more and more medical images are used in medical diagnosis. The sharp increase of medical image data makes the demand for medical image retrieval higher and higher. Under this background, people have proposed various solutions which can be broadly divided into following:(1) Text-Based Medical Image Retrieval tag each medical image tags like keywords and search for the corresponding images according to the keyword input, and then return the images matching keyword. This method is intuitive and in line with the people’s searching habits, but requires a lot of manual labeling of the images. (2) Content-Based Medical Image Retrieval extracting features of image input and images to be retrieved, and compare the similarity of features, then return images which meet the requirements of similarity. In this method, the process of image feature extraction is important. Because it will affect the final retrieval result depend on the feature is good or bad. At present, most of feature extraction methods are based on the analysis of image’s color, texture, shape and else which detemines the selection of feature extraction method should analyse the image first. So manual selection of feature extraction method is not universal.Under this background, this dissertation proposes a medical image retrieval method based on SAE feature extraction, thus to realize the automatic separation of image which including specified targets from MR images in the sequence. This dissertation has been supported by the open project of the key laboratory of high performance computing in Guangdong province(project number TH1528). The work of this dissertation is as follows:Firstly, analyze the research background, purpose and significance of this research. And the domestic and international research status of medical image retrieval system is introduced.Secondly, in the basis of analysis of medical image retrieval framework and visualized features of image, describes the PCA feature extraction principle and algorithm, the construction process of the AE feature extraction principle and algorithm and SAE, algorithm of SVM classifier, BP neural network classifier algorithm.Thirdly, present the medical image retrieval model based on SAE feature extraction. Then propose an improved SAE algorithm which adding random masking noise to the input on the basis of traditional SAE, so the auto encoder should learn to remove the noise and reconstruction the pure one in training process. Therefore, it forces the encoder to study the expression of input data is more robust. Finally, the improved SAE algorithm is used in the medical image retrieval model based on SAE feature extraction, and the training and construction steps of the model are described in detail.Fourthly, the medical image retrieval model based on SAE feature extraction is carried out and the experimental results are analyzed. In this dissertation, the trained model and another retrieval model which based on PCA feature extraction and SVM classification are used to test experiment on MR medical image database. The experimental results show that the recognition time of the two methods is 1.31s and 21.56s respectively, and the total retrieval accuracy is 93.66% and 89.53% respectively. The experimental results show that the SAE feature extraction method can achieve very good results for image retrieval.