Research On Multimodal Brain Image Analysis Based On Transfer Learning And Their Applications

Early diagnosis and intervention of brain disease has been of great importance for its therapy and studies of drug treatment in the clinic. Using brain image detection is widely used for early diagnosis of brain disease. Recently, machine learning methods are widely used for multimodal brain image analysis. However, most of the present machine learning methods for multimodal brain image analysis focus on the model design that use single learning domain data with supervised learning. In machine learning community, transfer learning has been developed to better deal with the problem involving relevant learning domain data. In this dissertation, we proposed a series of multimodal classification and regression methods based transfer learning for brain image analysis. In this dissertation, the main innovative research works are as following:(1) To effectively utilize knowledge from relevant learning domain data, a multimodal domain transfer learning classification model is proposed. Specifically, at first, a domain transfer feature selection(DTFS) component that selects the most informative feature-subset from both target domain and relevant learning domains; and then a domain transfer sample selection(DTSS) component that selects the most informative sample-subset from the same target and relevant learning domains; in the end, a domain transfer support vector machine(DTSVM) classification component that fuses the selected features and samples.(2) To effectively utilize knowledge from relevant learning domains and unlabeled data, a multimodal manifold-regularized transfer learning(M2TL) model is proposed. Specifically, the first one is a kernel-based maximum mean discrepancy criterion, which helps eliminate the potential negative effect induced by the distributional difference between the relevant learning domains and target domains. The second one is a semi-supervised multimodal manifold-regularized least squares classification method, where the target-domain samples, the relevant learning domain samples, and the unlabeled samples can be jointly used for training our classifier. Furthermore, with the integration of a group sparsity constraint into our objective function, the proposed M2 TL has a capability of selecting the informative samples to build a robust classifier.(3) To accurately evaluate the stage of brain disease and predict future progression, a semi-supervised multimodal relevance vector regression based on transfer learning(SM-RVR) is proposed, which jointly utilize multimodal features from brain images and biological biomarkers. According to the technique of multi-kernel learning, we proposed a kernel-fusion method based on multimodal features from brain images and biological biomarkers, which can be embedded it directly in the relevance vector regression, namely Multimodal Relevance Vector Regression(M-RVR). And then, we proposed a multimodal k-nearest neighbors regression method for predicting the clinical scores of unlabeled data. The next one is an M-RVR based recursive sample selection to select the most informative unlabeled data. The last one is training the M-RVR model to predict the clinical scores for the new test subjects.(4) To effectively utilize knowledge from relevant learning domain and multi-label data, a multi-domain multi-label(MDML) sparse feature learning based on sparse group Lasso is proposed, which can jointly select features from multi-domain and multi-label learning tasks. Specifically, the first one is multi-label sparse group Lasso feature learning model for fusing class label and multiple clinical scores. The second one is the MDML sparse feature learning model based on multi-label sparse group Lasso according to the idea of transfer learning. Then, all selected features are used for classification and regression with Multimodal Support Vector Machine(M-SVM) and M-RVR methods, respectively.All proposed method are trained and tested using Magnetic Resonance Imaging(MRI), Positron Emission Tomography(PET), and biological cerebrospinal fluid(CSF) markers data of subjects from the Alzheimer’s Disease Neuroimaging Initiative(ADNI).