Study Of Sparse Feature Learning Method In MRI Image Analysis

Recently, MRI image analysis has been widely used in studying brain structure and function, and computer-aided diagnosis methods for neural diseases. On the other hand, as the development of artificial intelligence, machine learning techniques, especially sparse feature learning, has widely used in MRI image analysis, playing an important role in building classification and prediction models. Thus, developing new methods for MRI image analysis is crucial for deeply mining information of MRI images and boosting the development of brain science study and computer-aided diagnosis. In machine learning-based MRI image analysis, we are often confronted with the problem that the number of samples is limited while the feature dimension is quite high, thus resulting in overfitting, noise and redundant features, which might severely degrade the performance of models. Sparse feaure learning can circumvent the difficulties mentioned above, and has been employed in signal processing, pattern recognition and computer vision. In this study, we are dedicated to developing new methods for MRI image analysis via desgning new regularizors for cost function for selection the most discriminative features, thus improving the performance of models in classification and prediction. Sparse feature learning methods in this study include single task sparse feature leanring: sparse Bayesian learning and 1L norm based sparse learning, as well as multi-task sparse feature learning: group Lasso, dirty model and sparse group Lasso. At the same time, we developed new algorithms based on methods mentioned above, and obtained excellent performance. Generally, this study includes five parts:1. Single voxel analysis tends to neglect correlations between neigburing voxels. We built a multi-voxel pattern analysis based learning model to study primary visual cortex decoding stimulated by spatial visua stimuli. Further, we proposed a multi-class decoder to combine feature selection and brain decoding. Specifically, the proposed model is capable of performing decoding, and at the same time selecting the most relevant features. The experimental results showed that 9 most relevant voxels were selected from the total of 2000 voxels, and the decoding accuracy achieved 91.6% via using the selected voxels. Meanwhile, we mapped the selected voxels into original brain space, and testified from another aspect the retinotopic mapping of primary visual cortex.2. As a pioneering work, we proposed a 1L norm and structural MRI images based method to diagnose Moyamoya disease. First, we extracted cotical thickness features from MRI images, obtaining approximately 20,000 features. Then, we built three types of 1L norm based feature selection models including Lasso, elastic net and L1-logistic regression, and significantly reduced feature dimensions with the methods above. With selected features, we trained a support vector machine(SVM) to diagnose the disease. Experimental results showed that proposed methods obtained satisfactory results, among which elastic net perfomed the best with the diagnosis accuracy of 82.36% and the area under ROC of 0.833. This result is significantly proior to the method without feature selection(diagnosis accuracy of 71.72%, and area under ROC of 0.787). Compard with conventional angiography based method, the proposd method is simple and non-invasive, thus can be used as a potential techinique for regular diagnosis.3. We built a multi-kerel support vector regession based IQ estimation model using strucral MRI images form children(6 years old to 15 years old), and proposed extended dirty model for multi-tasl feature selection. First, we extracted gray matter/white matter features from structural MRI images. Regarding feature selection for gray matter/white matter as a task, respectively, the proposed extended dirty model can be used. With selected features, we calculated kernel functions for gray matter/white matter, respectively, and then sent them to multi-kernel SVR to estimate IQ scores. The experimental results showed that correlation between estimated IQ scores by our proposed method and real IQ scores is 0.718, and the corresponding root-mean-square error is 8.695. Also, we selected 15 brain areas via the proposed method and found that, by referring to previous literatures, most of selected brain areas are associated with coginition and memory. The significance of the proposed IQ estimation model is that by predicting IQ scores of babies, we are able to design tailored plans for their early education.4. Regarding conventional functional MRI brain decoding needs to build a decoder for each subject separately, in this study, we proposed a multi-subject bran decoding framewok. Conventaionl decoding method uses voxels as features, whose active patterns have significant variability between subjects, thus difficul to build a single decoding model for all subjects. To solve this problem, we proposed a hierarchical model where voxels were treated as low level features and used to learn more robust high level features by SVM. Then, all high level features were sent to extended dirty model for feature selection. With selected high level features, a single decoding model can be built for all subjects. The proposed framework was validated by functional MRI images stimulated by 2D/3D visul stimuli, with classification accuracy achieving 89.4%, significantly prior to the method without feature selection. This method provides a potential method for study neural science of inter-subject.5. Conventional diagnosis method for autism depends on behavior-based scores. However, it is difficult to use a single score to diagnose such disease. In this study, we proposed to diagnose autism using structural MRI images and machine learning techniques. Also, we proposed a novel Canonical graph matching sparse group lasso algorithm for multi-task feature selection. First, we extracted gray matter/white matter, and mapped these features onto a new canonical space. We treated canonical feature selection for label and SRS_TOTAL score as two tasks, respectively, and then used the proposed method to select the most discriminative canonical features, with selected canonical features sent to SVM for classification. The experimental results showed that our proposed method achieved diagnosis accuracy of 75.4%, and area under ROC of 0.804, significantly prior to both competing methods using original gray matter/white matter features and other state-of-the-art methods using canonical features. The proposed method provides an alterative for assisting clinical physicians in autism diagnosis.