Research On Feature Extraction And Its Application Of Medical Image

With the development of computer vision and image processing technology,the medical image processing and analysis technology has become an important part of the computer-aided diagnosis.As the digital description method of image content and key attributes,image features play an important role in all aspects of image processing and analysis.However,thanks to the unique complexity in medical image,expected results are difficult to obtain through the traditional method of image feature extraction,which can't meet the needs of medical image processing and analysis of the practical application.This thesis is about the method of medical image extraction,researching on the application of image features in different cases such as classification,registration and segmentation of medical images,and extracting suitable image features according to the accuracy,stability and efficiency.Simultaneously,we improve our program according to the actual application demand.The original contribution of this thesis is mainly represented in the following four aspects:1.This thesis presents a medical image classification method based on combing two-dimensional texture features with three-dimensional morphological features.Aiming at the problem of lack of image feature discrimination and poor stability in a single type of medical image classification,a medical image classification method based on multi-type and multi-dimensional feature fusion is proposed.First,using two texture feature extraction methods: use the gray-level co-occurrence matrix and Gabor filter to extract the graphical features of brain MRI images on two-dimensional scale;Second,use morphological analysis method based on voxel to extract different biological features of different brain tissue volume and probability density,etc.in brain MRI images on three dimensions.Eventually,fusing these extracted features,and classify brain MRI images by SVM based on radial basis kernel.This method combines the low-level features and high-level features of medical images,and can describe and identify the image content from different angles simultaneously.The real brain MRI images segmentation experiment based on ADNI subset proves that this method can be used to detect and diagnosis the brain MRI image samples of patients with Alzheimer's disease and patients with mild cognitive impairment.Moreover,it has the ability to obtain good detection rate and accuracy.2.This thesis presents a two-stage medical image feature selection method based on improved SVM-RFEA two-stage optimal feature subset selection method is proposed for the problem of high feature dimension and redundant redundancy in medical image classification task based on multi-feature fusion.Firstly,the support vector machine recursive feature elimination method(SVM-RFE)is used to evaluate and order the initial features.Second,the correlation sequence between the features is obtained by using the covariance matrix.Then,through the sequence forward selection method(SFS),the optimal feature set is extracted under the feature recognition sequence and the correlation sequence,simultaneously search the optimal feature subset by two-sample detection.Finally,disentangle the image by the optimized feature subset.The real brain MRI images segmentation experiment based on ADNI subset proves that this method can effectively improve the detection rate and accuracy of classification algorithm and also reduce computational complexity.3.This thesis presents a medical image matching method that combines SIFT feature and 3D gray featuresAn improving method that combines the unique three-dimensional gray feature in medical image is proposed to the problem of matching mistakes by employing traditional SIFT feature descriptor in feature point matching task.First,take advantages of key features of scale-invariant feature transform(SIFT)to extract key features of the registered image and form the initial matching point pair;second,extract the initial matching point pair and its three-dimensional grayscale features of adjacent tomographic images in the stereoscopic brain image;last,the three-dimensional gray-scale feature is used to filter and adjust the initial matching point pairs,and then,the exacted feature matching results are obtained.Experiments have been carried out by extracting the real brain CT images from the same subjects at different times.The results showed that the method can obtain more accurate feature matching points,thus improving the precision of medical image registration.4.This thesis presents a medical image segmentation method based on deep neural networkAn interest region segmentation method base on deep learning frame is proposed for the problem of difficulty in discovering and extracting the useful features.First,using the unlabeled image block sample to build the depth learning network model by stacking the noise automatic reduction encoder,automatically learning and extracting the image features;then,using the labeled samples to fine-tune the depth learning network to enable the classification ability;then,we use the above model to classify the segmented samples to detect the initial segmentation region of the brain tumor;finally,the initial segmentation results are optimized by threshold segmentation and morphological methods to achieve accurate segmentation results.The results of segmentation experiment based on real brain tumor patients prove that this method can effectively improve both the accuracy and sensitivity of classification algorithm.Moreover,it can greatly improve running speed comparing to traditional machine learning methods.