Research On The Classification Of Early Mild Cognitive Impairment Based On Multi-task Learning

Detection of Alzheimer’s disease(AD)in its early stages(i.e.,mild cognitive impairment(MCI))is important because it can delay or prevent its progression to AD.Brain connection networks deduced from medical imaging data have been commonly used to identify patients with MCI from normal controls(NC).However,most of the existing methods are still limited in performance,and these methods are mainly developed by using only a single modal data for classification.In fact,it has been proven that multi-modal data can reflect brain information from different aspects,and providing complementary information is more conducive to the diagnosis and classification of diseases.At the same time,because structural and functional brain networks characterize the interaction of brain information from different levels of brain connection,the combined analysis of the two is more conducive to the diagnosis and prediction of disease progression in AD.Typical brain network analysis and disease diagnosis model framework mainly includes brain network construction,feature learning and classification prediction.That is,the structural or functional brain network is estimated first,and then the feature learning is performed from the estimated brain network.Finally,the learned features are used to build a model to complete the classification of brain diseases,and realize the diagnosis and prediction of samples.Based on this,this paper proposes two models for automatic diagnosis and prediction of early stage of MCI.First,low-rank self-calibrated brain network estimation and joint non-convex multi-task learning framework.Specifically,we first develop a new functional brain network estimation method.The introduction of a data quality indicator for selfcalibrated can improve the data quality while completing the brain network estimation,and perform correlation analysis in combination with the low-rank module structure.Second,functional and structural neuroimaging patterns are integrated into our multitask learning model to select discriminative and informative features for early AD analysis.Different modalities perform specific different classification tasks,and determine the similarities and differences between multiple tasks through joint learning to determine the most discriminative features.This learning process is completed by a non-convex regularizer,which effectively reduces the penalty deviation of the trace norm and approximates the original rank minimization problem.Finally,the most relevant features of the disease are fed to a support vector machine classifier for early AD recognition.The diagnostic accuracy of this method for 6 different classification tasks including AD vs SMC,NC vs EMCI,NC vs LMCI,SMC vs EMCI,SMC vs LMCI and EMCI vs LMCI are 82.95%,85.23%,87.80%,84.09%,90.24% and 81.71%,indicating that the proposed method can be effectively applied to the diagnosis and analysis of early MCI.Second,it integrates functional connections and structural connections for autoweighted centralized multi-task learning framework.Specifically,a functional brain network is constructed by sparse low-rank brain network estimation method,and structural brain network is constructed using fiber bundle tracking.Subsequently,we use multi-task learning methods to integrate the features of functional and structural connections.The importance of each task and the balance between the two modalities are automatically learned.By integrating functional and structural information,the most discriminative features of the disease can be obtained for diagnosis.This method has achieved an average accuracy of 84.80% / 89.12% / 87.82% for different classifications of NC / SCD / MCI.A large number of experimental results show that our algorithm has advantages.Extensive experiments were performed to prove the effectiveness of the proposed method on public datasets.This paper uses data from two modalities,including functional magnetic resonance imaging(fMRI)and diffusion tensor imaging(DTI).The leave-one-out(LOO)cross-validation method was used to evaluate the classification performance of the proposed method.The method proposed in this paper has good performance and is superior to other traditional algorithms.In addition,the proposed method identifies disease-related brain regions and connections.These results are consistent with current clinical findings and add new findings for disease detection and future medical analysis.