Research And Application Of Medical Image Recognition Based On Multi-feature Fusion

With the ubiquity of medical image acquisition equipment(such as X-rays, CT, MRI), Medical images turn into the Indispensable Carrier of record and preserve patient physiological disease information. On one hand, as time goes by, Numerous of medical image are preserved, it's an additional burden for doctors, and the complexity of some diseased organs makes giving an objective judgment be difficult just relying on the naked eye. On the other hand, with the further development of computer performance and pattern recognition technology, it is a step forward for medical image recognition by computer technology. However, since the medical image have rich information, including complicated structure organs, it's difficult to extract features and recognize images. Single feature only contains certain aspects of information, lacking comprehensiveness. So how to apply feature fusion technology to medical image recognition is an important research on Interdisciplinary of the computer science and medical field.This paper mainly studies the medical image feature extraction and multi-feature fusion algorithm and its application in the identification and classification, including of the following detailed innovations and contributions:1. We propose a fast SURF extraction algorithm based on hybrid filter Haar convolution template. By improving Haar wavelet response convolution template, under the condition of not relying on additional resources and loss characteristics of discrimination, the SURF feature extraction time saves about 27%.2. We improve multi-feature fusion based on the kernel function of Canonical correlation analysis. Each of the canonical correlation coefficient is given a different weight information, making more notable between features. In the feature fusion and medical image recognition, increased by about 5% and 10%, compared to other mainstream fusion algorithm and single feature, respectively.3. We propose hybrid attributes similarity measurement for Spectral clustering. By introducing the concept of the local density and local scales, we make similarity matrix more agglomerated. It's simple and effective, dramatically improving the accuracy rate, compared to other mainstream spectral clustering algorithms.4. We propose a new multi-feature fusion algorithm based on visual dictionary learning. In terms of visual dictionary structure, unlike BOW ideas, we use a hybrid attributes similarity measurement to build visual word. By LR model and L1 regularization, each type of image can be expressed by a small number of visual discrimination word, and use the output of LR model to generate a marginalized kernel function. Experimental results show that the recognition rate has improved about 20% and 6%-10%, compared to single feature and SLF, WKCCA fusion algorithm, respectively, overall recognition rate of about 87%.