Multi-resolution 3D-HOG feature learning method for Alzheimer’s Disease diagnosis

https://doi.org/10.1016/j.cmpb.2021.106574Get rights and content

Highlights

  • A multi-resolution 3D-HOG feature extraction method.

  • Local and global texture describer of AD.

  • Histogram based wrapped feature selection method.

  • Detect distinct subareas of ROIs for AD identification.

Abstract

Background and Objective: Alzheimer’s Disease (AD) is a progressive irreversible neurodegeneration disease and thus timely identification is critical to delay its progression. Methods: In this work, we focus on the traditional branch to design discriminative feature extraction and selection strategies to achieve explainable AD identification. Specifically, a spatial pyramid based three-dimensional histogram of oriented gradient (3D-HOG) feature learning method is proposed. Both global and local texture changes are included in spatial pyramid 3D-HOG (SPHOG) features for comprehensive analysis. Then a modified wrapper-based feature selection algorithm is introduced to select the discriminative features for AD identification while reduce feature dimensions. Results: Discriminative SPHOG histograms with various resolutions are selected, which can represent the atrophy characteristics of cerebral cortex with promising performance. As subareas corresponding to selected histograms are consistent with clinical experience, explanatory is emphasized and illustrated with Hippocampus. Conclusion: Experimental results illustrate the effectiveness of the proposed method on feature learning based on samples obtained from common dataset and a clinical dataset. The proposed method will be useful for further medical analysis as its explanatory on other region-of-interests (ROIs) of the brain for early diagnosis of AD.

Introduction

Alzheimers Disease (AD) is a progressive neurodegenerative disease, which reflects anatomical atrophy or functional neurodegeneration of cerebral cortex. In recent years, some machine learning methods have been used to extract useful features from magnetic resonance imaging (MRI) scanned anatomical data to identify AD from Healthy Controls (HC). For feature extraction methods, three-dimension (3D) image based methods can effectively preserve the spatial feature information of MRI data compared with two-dimension (2D) image based methods. These methods directly or indirectly extract 3D features from MRI data using traditional feature extraction methods or deep learning methods respectively.

Most traditional feature extraction methods directly extract anatomical features of cerebral cortex from MRI images. Among them, the feature-based machine learning methods use the clinical parameters as features, such as the volume of gray matter, the cortical thickness, the mean curvature and area of cortical which are extracted by FreeSurfer image analysis suite [1]. Some image-based feature extraction methods extract features from different transform domain, such as texture-based Gabor transform method [2], multi-resolution-based discrete wavelet transform (DWT) method [3]. And there are other image-based methods, which directly extract image-based feature to describe the atrophy or shape changes of region-of-interests(ROIs) of the brain, such as ROIs-based sparse feature learning method [4], [5], local binary pattern (LBP) method [6] and histogram of oriented gradient (HOG) method [7], [8]. As an image gradient based feature extraction method, HOG method extracts image gradients within a region to reflect its edge gradient changes [9], which is used for region detection in medical images [10], [11]. Considering that there is volume atrophy and shape changes of cortical in AD, HOG method is also used to detect the local texture shape changes or gradient changes of the image for early diagnosis of AD. For example, Devvi et.al extracted HOG feature from three orthogonal of planes to describe the dynamic texture changes of MRI brain images  [7]. Zhu et.al proposed a new multi-view learning method to learn the mappings from the HOG feature space to the ROI feature space, which uses 3D-HOG features as local features to reflect small or subtle changes within brain [12]. And in [8], small scale HOG features are extracted from ROIs and used to quantify spatial gradients of 18F-FDG PET images for AD diagnosis. 3D-HOG features represent the local texture changes within a volume statistic of spatial gradient and overcome the information loss generated from 2D-HOG representation. However, there are two main limits for 3D-HOG method: (1) features with same scale only represent local visual features with the same resolution, which cannot represent the characteristics of the image comprehensively. (2) HOG features consist of numerous histograms and thus cover invalid information. Effective characteristic should be obtained in additional step for accurate classification.

Although there are exhaustive extracted features, irrelevant or redundant features may reduce the efficiency of learning algorithms, i.e. not all extracted features are useful for the classification problems. As discussed in [13] and [14], feature extraction usually encounter the so-called ’High Dimension, Low Sample Size (HDLSS)’ problem. In order to resolve this problem, subspace learning methods and feature selection methods are used to reduce the feature dimensions to choose discriminating features. Subspace learning methods include linear methods, such as principle component analysis (PCA) [4], [15], [16], linear discriminant analysis (LDA) [17], [18], and non-linear methods, such as multi-kernel methods [19], [20]. Feature selection methods generally choose discriminative feature subset for the following classification, which are divided into class-dependent methods and class-independent methods. Class-independent feature selection methods choose potential features while ignoring different classes. Class-dependent feature selection methods utilize different feature subsets to discriminate different classes and obtain better performance than class-independent feature selection [21], which can be further divided into filter approaches and wrapper approaches depending on whether classifier is used or not. Filter approaches utilize various feature importance ranking methods for feature selection and selected features are used for comprehensive classification. For example, Minimal-RedundancyMaximal-Relevancy Measure (mRMR) method [22] selects attributes with maximal relevance and the minimal redundancy based on calculating the mutual information. RELIEF method [23] is a weighted method which tends to minimize intra-class distance and maximize inter-class distance. Class Separability Measure (CSM) method [24] is proceeded by calculating the intra-class and inter-class ratio which is used to evaluate the contribution of each attribute. The wrapper based method [21] finds discriminative feature subsets for each class and then uses the class-dependent subset for final classification. In [25], we proposed a wrapper-based feature selection method to rank the feature importance, which can select most important feature parameters or featured ROIs for AD identification compared with other feature selection methods.

As the image-based classification methods, deep learning methods use the whole brain or ROIs as input of the network to extract 2D or 3D features from MRI data indirectly. There are different network structures introduced to extract discriminate features from original images, such as convolution neural network (CNN), Residual Network (ResNet) [26], a combination of encoder-decoder network [27], U-Net [28], DenseNet [29], and transfer learning strategy [30]. Although these methods have shown good performance on AD classification, it is difficult to present interpretability on the extracted features or classification results because they incorporate feature extraction and classifier learning into an unified framework [31], [32]. Furthermore, the training outcome is unsatisfactory as the lack of samples [33].

In this paper, we proposed a ROIs-based multi-resolution 3D-HOG feature learning method for AD identification. Some preliminary accounts of this study were presented in our early conference papers [25], [34]. The main contributions of this paper can be concluded as follows:

  • First, we proposed a multi-resolution 3D-HOG feature extraction method to describe local and global texture changes for AD identification, which can represent the characteristics of the image comprehensively compared with previous HOG-based approaches.

  • Second, we proposed a histogram based wrapped feature selection method, which can not only select discriminative histograms with promising performance, but also detect distinct subareas of ROIs for AD identification.

The remaining of this paper is organized as follows: Section 2 presents the proposed feature extraction and feature selection method, Section 3 presents the experimental results, followed by the conclusion of this paper in Section 4.

Section snippets

Methodology

Framework of the proposed method is shown in Fig. 1. First, data preprocessing with MRI T1-weighted input is used to generate ROI-based images for feature extraction and classification. Second, 3D-HOG from various scales is extracted as basic feature unit. Third, spatial pyramid HOG features are constructed with multi-scale 3D-HOG features for informative representation. Finally, feature selection techniques are introduced to search for discriminative features and further efficient

Experimental result and discussion

In the following experiments, 20 times Monte-Carlo simulations are carried out to: (a) illustrate the effectiveness of SPHOG; (b) illustrate the effectiveness of the proposed feature selection method; (c) compare with other traditional machine learning methods and deep learning methods; (d) illustrate the effectiveness of the selected features. (e) illustrate the effectiveness of our proposed methods on multi-region based identification.

65 AD subjects and 65 HC subjects (totally 130 samples)

Conclusions

In this paper, we proposed a novel feature learning method for AD identification. From spatial pyramid representation, the multi-resolution SPHOG features are constructed to distinguish the deformation characteristics of cerebral cortex comprehensively. With the proposed histogram based wrappered feature selection algorithm, the discriminative SPHOG features are selected and the feature dimensions are reduced. Experimental results show that the selected SPHOG features outperforms other 3D-HOG

Declaration of Competing Interest

The authors declare that we have no financial and personal relationships with other people or organizations that can inappropriately influence our work. There is no professional or other personal interest of any nature or kind in any product, service and/or company that could be construed as influencing the position presented in, or the review of, the manuscript entitled.

References (50)

  • H. Kilian et al.

    Adaptive fusion of texture-based grading for Alzheimer’s disease classification

    Computerized Medical Imaging and Graphics

    (2018)
  • Y. Liu et al.

    Compartmental sparse feature selection method for Alzheimer’s disease identification

    Conf Proc IEEE Eng Med Biol Soc

    (2017)
  • H. Bhasin et al.

    A combination of 3-d discrete wavelet transform and 3-d local binary pattern for classification of mild cognitive impairment

    BMC Med Inform Decis Mak

    (2020)
  • D. Sarwinda et al.

    3D-hog features-based classification using mri images to early diagnosis of Alzheimer’s disease

    17th IEEE/ACIS International Conference on Computer and Information Science, ICIS

    (2018)
  • X. Pan et al.

    Multiscale spatial gradient features for 18f-fdg pet image-guided diagnosis of Alzheimer’s disease

    Comput Methods Programs Biomed

    (2019)
  • N. Dalal et al.

    Histograms of oriented gradients for human detection

    IEEE Computer Society Conference on Computer Vision and Pattern Recognition

    (2005)
  • M. Erdt et al.

    Region detection in medical images using hog classifiers and a body landmark network

    (2013)
  • X. Zhu et al.

    Multi-view classification for identification of alzheimer’s disease

  • C. Chu et al.

    Does feature selection improve classification accuracy? impact of sample size and feature selection on classification using anatomical magnetic resonance images

    Neuroimage

    (2012)
  • Y.D. Zhang et al.

    Classification of Alzheimer’s disease based on structural magnetic resonance imaging by kernel support vector machine decision tree

    PROG ELECTROMAGN RES

    (2014)
  • E. Adeli-Mosabbeb et al.

    Robust feature-sample linear discriminant analysis for brain disorders diagnosis

    (2015)
  • Z. Wang et al.

    Classification of Alzheimer’s disease, mild cognitive impairment and normal control subjects using resting-state fmri based network connectivity analysis

    IEEE J Transl Eng Health Med

    (2018)
  • J. Peng et al.

    Structured sparsity regularized multiple kernel learning for Alzheimer’s disease diagnosis

    Pattern Recognit

    (2019)
  • H. Peng et al.

    Feature selection based on mutual information: criteria of max-dependency, max-relevance, and min-redundancy

    IEEE Trans Pattern Anal Mach Intell

    (2005)
  • K. Kira et al.

    The feature selection problem: traditional methods and a new algorithm

    Tenth National Conference on Artificial Intelligence

    (1992)
  • This work was supported by “National Natural Science Foundation of China (NSFC)” ( No. 61971106, 61603077, 41776204), “Fundamental Research Funds for the Central Universities”

    View full text