Multi-modal discriminative dictionary learning for Alzheimer's disease and mild cognitive impairment
Introduction
Studies of Alzheimer's disease (AD) and mild cognitive impairment (MCI) have explored varied neuroimaging modalities with promising results. These include structural Magnetic Resonance Imaging (sMRI) [41], [52], functional MRI (fMRI) [15], [32], Flourodeoxyglucose Positron Emission Tomography (FDG-PET) [23], and amyloid PETs such as Pittsburgh compound B (PiB-PET) [51], florbetapir [27], and flutemetamol [36] PETs. As each single modality offers specific information about MCI or AD, combining the complementary information from different modalities might enhance understanding of AD and MCI [10], [29], [43], [48].
sMRI provides structural information about the cerebrum and has proved that brain regions such as the hippocampus and parahippocampus are most suited for differentiation between MCI and normal controls (NC) [6], [11], [22], [41]. FDG-PET measures glucose metabolism and has indicated that brain regions such as the superior frontal gyrus, and middle cingulate cortex discriminate well between MCI and NC [9], [18], [23]. Amyloid-PET non-invasively measures the accumulation of amyloid in the brain, and it has suggested that brain regions such as the posterior cingulate and lateral temporal cortices are affected more in MCI than the NC [4], [30]. These recent studies demonstrate that each brain-imaging technique can provide specific views about brain function or structure [3]. In other words, biomarkers from these modalities offer different and potentially complementary information about various aspects of a given disease process [2], [5], [27]. Indeed, multi-modality neuroimaging has been viewed as a research method in neuroscience [31], [53].
Numerous studies have reported various ways of combining multi-modality data for efficient classification [8], [10], [16], [33], [34], [42], [50] and better differentiation of patients with AD or MCI from cognitively healthy individuals. For example, a weighted multiple kernel learning (MKL) model has been applied to combine cerebrospinal fluid (CSF), MRI, and PET to produce more powerful classifiers of MCI [16]. A linear weighted random forest (RF) model could combine different modalities and discriminate AD or MCI from NC effectively [8]. To utilize both the basic simple features and complex latent representation, multi-kernel SVM learning has been applied [33], [34], [42]. These studies demonstrate that the weighted combination approach is a simple yet effective way to integrate information from multi-modality data.
Recently, weighted multi-modality sparse representation-based classification (mSRC) has been introduced among neuroimaging communities and has demonstrated its feasibility and effectiveness in discriminating AD or MCI from NC by using data from sMRI, FDG-PET, and florbetapir-PET [44]. However, there are some weaknesses in mSRC as it is based on the SRC method that uses all the training data as a dictionary. When faced with large training sets, the sparse representation of a dictionary might be computationally time-consuming [13]. In addition, when training samples are not very representative, the classification accuracy would be affected by the dictionary [46].
A recently introduced method in pattern recognition and machine learning by Xu et al. considered supervised within-class-similarity discriminative dictionary learning (SCDDL) a robust and efficient method for facial recognition [45]. To improve the accuracy of facial recognition, SCDDL incorporated the term “within-class-similarity” for representation coefficients into the objective function of dictionary-learning. This is a combination of Fisher Discrimination Dictionary Learning (FDDL) [47] and Discriminative K-SVD (D-KSVD) [49] or Label Consistent K-SVD (LC-KSVD) [12]. wmSRC in combination with a linear classification error term has been introduced. This aims to derive a more compact dictionary as compared with SRC [45]. Although its first application was in facial-recognition (2-dimensional data), we believe that SCDDL has potential promise in the field of neuroimaging data, especially in multi-modal data for differentiation of AD or MCI from NC.
The contributions of this study are as follows. First, the SCDDL method was extended from single to multiple modalities based on weighted combination (mSCDDL), and was examined with regard to its robustness and accuracy of differentiating NC from MCI or AD. For this study, the multi-modal data used in SCDDL and mSCDDL were sMRI, FDG-PET, and florbetapir-PET. Second, the mSCDDL method was compared with other state-of-the-art multi-modality classification algorithms for performance in differentiating NC from MCI or AD.
Section snippets
Participants
The datasets used in this study were downloaded from the Alzheimer's Disease Neuroimaging Initiative (ADNI) (http://www.loni.ucla.edu/ADNI/). A $60 million 5-year project, the ADNI was launched in 2003 by the National Institute on Aging (NIA), National Institute of Biomedical Imaging and Bioengineering (NIBIB), Food and Drug Administration (FDA), private pharmaceutical companies, and non-profit organizations.
According to ADNI protocols, the severity of cognitive impairment was assessed using
Algorithm
All the features extracted from three modalities of data (sMRI, FDG-PET, and florbetapir-PET) were included for the differentiation of AD and MCI from NC. The supervised within-class-SCDDL was presented for the first time in this section and was applied to neuroimaging data to classify MCI and AD from NC. Further, the extended multi-modality framework based on SCDDL, termed as multi-modality SCDDL (mSCDDL), was applied to the combined multi-modality data to differentiate NC from MCI and AD
Comparison with single-modality SCDDL
To compare the results more easily, the dictionary size was set as 20 atoms for both SCDDL and mSCDDL. The performances of single-modality SCDDL (SCDDL-sMRI, SCDDL-FDG-PET, and SCDDL- florbetapir-PET) and multi-modality mSCDDL (sMRI + FDG-PET + florbetapir-PET) were evaluated. The multi-modality mSCDDL achieved higher accuracy in classifying MCI or AD from NC than all the single-modality methods as shown in Fig. 1 and Table 2.
For classifying MCI from NC, the mSCDDL achieved an accuracy of
Discussion
In this study, a multi-modality classification method, mSCDDL was extended from single modality and compared with other state-of-the-art multi-modality methods (MKL, JRC, and mSRC) to identify AD and MCI. Three modalities, namely sMRI, FDG-PET, and florbetapir-PET, were used. The results revealed the effectiveness of mSCDDL for differentiation between AD and MCI (97.36% for AD and 77.66% for MCI). And the mSCDDL method was compared with other state-of-the-art multi-modality classification
Limitations
This study had several limitations. First, many other data sources are useful for AD or MCI classification in addition to sMRI, FDG-PET, and florbetapir-PET, such as cerebrospinal fluid (CSF) [28], [35], [50]. Second, only the weighted combination method was used in this study for the multi-modality analysis. Additional studies could attempt to expand SCDDL to multi-modal ones in the framework of multi-kernel learning. Third, this study did not include the cognition information in the
Conclusions
This study proposed a multi-modality-supervised within-class-similar discriminative dictionary learning classifier called “mSCDDL” to combine the multi-modality features (sMRI, FDG-PET, and florbetapir-PET) for differentiating AD and MCI from NC. The results suggest that the mSCDDL procedure is a promising tool for classification especially in helping to diagnose diseases using neuroimaging data.
Funding information
This work was supported by the Funds for International Cooperation and Exchange of the National Natural Science Foundation of China [grant number 61210001], the General Program of National Natural Science Foundation of China [grant number 61571047], and the Fundamental Research Funds for the Central Universities [grant number 2017EYT36].
Acknowledgements
The data set used in preparation of this paper was obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database (adni.loni.ucla.edu). As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data but did not participate in the analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.ucla.edu/wcontent/uploads/how_to_apply/ADNI_Acknowledgement_ist.pdf.
Declaration of interest
(1) There are no actual or potential conflicts of interest.
(2) There is no author's institution has contracts relating to this research through which it or any other organization may stand to gain financially now or in the future.
(3) All of the authors could be seen as involving a financial interest in this work.
Submission declaration and verification
The work described has not been published previously, and will not be submitted elsewhere while under consideration at Pattern Recognition.
All authors have reviewed the contents of the manuscript being submitted, approve of its contents and validate the accuracy of the data.
References (56)
A fast diffeomorphic image registration algorithm
Neuroimage
(2007)Correlations between FDG PET glucose uptake-MRI graymatter volume scores and apolipoprotein E (4 gene dose incognitively normal adults: a cross-validation study usingvoxel-based multi-modal partial least squares
Neuroimage
(2012)Spatialpatterns of brain atrophy in MCI patients, identified viahigh-dimensional pattern classification, predict subsequentcognitive decline
Neuroimage
(2008)“Mini-mental state”: a practical method for grading the cognitive state of patients for the clinician
J. Psychiatr. Res.
(1975)- et al.
Categorical and correlational analyses of baseline fluorodeoxyglucose positron emission tomography images from the Alzheimer's disease neuroimaging initiative (ADNI)
Neuroimage
(2009) Single domain amnestic MCI: a multiple cognitive domains fMRI investigation
Neurobiol. Ag.
(2011)Connectivity analysis of normal and mild cognitive impairment patients based on FDG and PiB-PET images
Neurosci. Res.
(2015)Hierarchical feature representation and multimodal fusion with deep learning for AD/MCI diagnosis
Neuroimage
(2014)[18F] flutemetamol amyloid positron emission tomography in preclinical and symptomatic Alzheimer's disease: specific detection of advanced phases of amyloid-β pathology
Alzheim. Dement.
(2015)Fast and robust parameter estimation for statistical partial volume models in brain MRI
Neuroimage
(2004)
Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain
Neuroimage
Multi-modal imaging predicts memory performance in normal aging and cognitive decline
Neurobiol. Ag.
Enriched white matter connectivity networks for accurate identification of MCI patients
Neuroimage
Identification of MCI individuals using structural and functional connectivity networks
Neuroimage
Combining MRI and CSF measures for classification of Alzheimer's disease and prediction of mild cognitive impairment conversion
Neuroimage
Multi-modality sparse representation-based classification for Alzheimer's disease and mild cognitive impairment
Comput. Methods. Programs. Biomed.
Multi-source feature learning for joint analysis of incomplete multiple heterogeneous neuroimaging data
Neuroimage
Multimodal classification of Alzheimer's disease and mild cognitive impairment
Neuroimage
A novel matrix-similarity based loss function for joint regression and classification in AD diagnosis
Neuroimage
Multimodal Analysis in Normal Aging, MildCognitive Impairment, and Alzheimer's Disease: Group Differentiation, Baseline Cognition and Prediction of Future Cognitive Decline
A method for multitask fMRI data fusion applied to schizophrenia
Human Brain Mapp.
Using PET with 18F-AV-45 (florbetapir) to quantify brain amyloid load in a clinical environment
Eur. J. Nucl. Med. Mole. Imag.
Random forest-based similarity measures for multi-modal classification of Alzheimer's disease
Neuroimage
Glucose metabolism during resting state reveals abnormal brain networks organization in the Alzheimer's disease and mild cognitive impairment
PLoS One
Predictive markersfor AD in a multi-modality framework: an analysis of MCI progression in the ADNI population
Neuroimage
Prediction of AD with MRI-based hippocampal volume inmild cognitive impairment
Neurology
Label consistent K-SVD: learning a discriminative dictionary for recognition
IEEE Trans. Pattern Anal. Mach. Intel.
Comparing predictors of conversion and decline inmild cognitive impairment
Neurology
Cited by (30)
A machine learning approach to screen for preclinical Alzheimer's disease
2021, Neurobiology of AgingCitation Excerpt :Thus, quantitative EEG seems to be a promising tool to identify preclinical AD subjects, but further studies are needed to assess the individual diagnostic accuracy. Furthermore, combining multimodal biomarkers could help in the early diagnosis of AD (Frölich et al., 2017; Gupta et al., 2019; Li et al., 2017; Ritter et al., 2015). Previous studies have proposed multimodal models to predict brain amyloidosis in cognitively normal individuals (Ansart et al., 2020; Insel et al., 2016; Mielke et al., 2012; ten Kate et al., 2018).
CMC: A consensus multi-view clustering model for predicting Alzheimer's disease progression
2021, Computer Methods and Programs in BiomedicineCitation Excerpt :AD using multi-model data has also been studied by some researchers. Li et al. considered a multi-modal supervised within-class-similarity distinctive dictionary learning method based on the weighted combination for AD diagnosis by using the neuroimaging data sets[21]. Tong et al. proposed a multi-modality classification model with nonlinear graph fusion to utilize the complementary information among different modal data of AD [48].
Hi-GCN: A hierarchical graph convolution network for graph embedding learning of brain network and brain disorders prediction
2020, Computers in Biology and MedicineCitation Excerpt :Brain disorders such as Alzheimer's disease (AD) [1–4] and Autism spectrum disorder (ASD) [5,6] are considered in terms of disruptions of the normal-range operation of brain functions.
Predicting the progression of mild cognitive impairment to Alzheimer's disease by longitudinal magnetic resonance imaging-based dictionary learning
2020, Clinical NeurophysiologyCitation Excerpt :By using the label information available in the training set, supervised dictionary learning methods exploit the discriminating information for classification and deliver good classification performance in natural scene classification(Wang and Kong, 2014) and face recognition(Yang et al., 2014) tasks. Recently, dictionary learning schemes have also been utilized in medical applications and can be used to directly extract features from a sparse matrix without the segmentation of ROIs(Vu et al., 2016; Diamant et al., 2017; Li et al., 2017). In this paper, we developed an objective and automated prediction system using patch-based dictionary learning on a longitudinal structural MRI to automatically obtain the subtle gray matter (GM) density differences between patients with pMCI and sMCI.
Use of artificial intelligence in Alzheimer’s disease detection
2020, Artificial Intelligence in Precision Health: From Concept to ApplicationsAutomated atrophy assessment for Alzheimer's disease diagnosis from brain MRI images
2019, Magnetic Resonance ImagingCitation Excerpt :Worst accuracy performance results are described in [44] with the highest feature dimension (4761 features/image) also, thus results in high computational complexity. The techniques described in [46–48] show amended results in brain MR image classification, with lower feature vector dimension (19), but these schemes use various complex weight optimization techniques, which themselves require high computational complexity. Feature vector dimension (7), used in [49], is equal to our proposed method (7), but it is obvious from the results of Table 3, that the method described in [49] is less effectual and general than the proposed scheme, in terms of classification accuracy whereas, our proposed system only requires a feature vector of dimension 7, with the highest retrieval accuracies.