Alzheimer’s disease (AD) is associated with abnormal functioning of the nervous system and usually appears in people over the age of 60. Patients are often accompanied by memory loss and cognitive decline and other problems, which not only seriously harms the physical and mental health of patients, but also brings a heavy burden on families and society [1]. It is estimated that in 30 years there will be 134.6 million cases of AD worldwide [2]. However, the exact cause of AD is still unknown and existing targeted drugs can only reduce symptoms or delay its progression. Therefore, revealing the brain changes caused by the disease is crucial to explore the underlying cause of AD [3].

Precise diagnosis of AD helps patients improve future quality of life, including early prevention and optimal treatment [4]. With the fast development of artificial intelligence and medical imaging technology, computer-aided diagnosis provides sufficient evidence of accuracy to distinguish AD from healthy control (HC) [5, 6, 7, 8, 9, 10]. Traditionally, AD has been thought of as a disease of gray matter (GM) damage, while the effects of white matter (WM) have generally been thought of as damage secondary to GM [11]. Although there is growing concern about WM damage in AD, our knowledge is still limited when compared to GM atrophy and other biomarkers. In particular, a review illustrates different main research point of penetration for how WM injury leads to AD [12]. One piece of evidence is to study WM degeneration and demyelination as the important pathophysiological features of AD at the microstructural level [13, 14]. Another piece of evidence comes from neuroimaging, which have the remarkable advantage of being able to noninvasively observe morphologic changes in patients’ brains. Despite the study is focused on WM microstructures, it’s important to highlight that diffusion tensor imaging (DTI) can be used to investigate even microstructural changes in GM (not only WM) [15, 16]. With the development of DTI technology, it can be used to show the direction of fasciculus in the WM of the brain, which is the only non-invasive imaging method that can show the fasciculus in vivo [17].

There are some common metrics that can reflect the brain microstructure in DTI, such as fractional anisotropy (FA), mean diffusivity (MD), axial diffusivity (DA) and radial diffusivity (RD). FA represents the directivity of water molecular dispersion and can reflect the maximum possible arrangement direction of WM tracts. The FA value is higher in WM, close to 1, while it is close to 0 in cerebrospinal fluid. MD represents the overall dispersion of water molecules. DA represents the degree of dispersion of water molecules along the main direction. RD represents the dispersion of water molecules in the other two directions [18, 19]. For further understand the pathological mechanisms of WM tracts’ change, tract-based spatial statistics (TBSS) method has been applied to research the microstructural of WM [20]. In recent years, machine learning-based neuroimaging technology for AD diagnosis and disease development has become a research hotspot [21, 22, 23, 24]. As a widely used supervised learning method, support vector machine (SVM) shows good advantages in solving small sample, nonlinear and high-dimensional pattern recognition problems [25, 26, 27, 28].

Here, we aimed to analyze research microstructural difference of WM and predict the accuracy between ADs and HCs. Therefore, the TBSS and SVM method will be used in our study. Specifically, we combined with different kind of DTI metric together for improving classification performance and rank the importance of the WM fiber tract.

In our study, two aspects were mainly researched: statistical analysis of DTI data and classification. The subjects were firstly preprocessed using PANDA tool. Then the statistical analysis was operated by TBSS and classification process was conducted by SVM method.

Relevant literatures showed that the association fiber and limbic system were the most reported abnormal regions in the WM tracts of AD [46, 47, 48]. Cingulate was the association fiber between cingulate gyrus and other brain structures. Its integrity might directly relate to the emotion and cognitive function in AD patients. In our study, FA value of the right cingulate gyrus in ADs obviously decreased compared to HCs that was agreed with previous researches [49, 50]. In addition, association fiber contacted with part cortex of the ipsilateral hemisphere, FA value of bilateral sagittal stratum, superior fronto-occipital fasciculus and superior longitudinal fasciculus in ADs decreased compared with HCs that was agreed with Teipel’s research [51]. Correlational study found superior fronto-occipital fasciculus influenced visual spatial processing and memory function [52]. The damage of superior longitudinal fasciculus might involve spatial working memory and linguistic function [53]. The decrease of these functions was reflected on ADs than related to our research results. Optic radiation was the central neurons of visual pathway so that its lesion would lead to defect of field vision. In our study, decrease of FA value of posterior thalamic radiation in ADs hinted the damage of visual performance [54]. Wang et al. [55] based on TBSS with multi-parameter found that bilateral hippocampus gyrus existed obvious abnormal in ADs and patients with mild cognitive impairment (MCIs), including the decrease of FA value and increase of RD value, especially the right hemisphere in which was most significant. While our study found the difference just exists in the right hippocampus gyrus. MD value increased when the tissue damage. Increased MD and decreased FA were found in the corpus callosum that was agreed with previous research [56]. The corpus callosum was a bundle of fibers connecting the right and left hemispheres of the brain. On the basis of previous researches, the anterior part of the corpus callosum was connected to the prefrontal cortex and was associated with the sense of motivation [57].

Therefore, our results suggested that communication disorders between brain structures might be related to apathy symptoms of ADs [58, 59]. Moreover, the particular pattern of association between severity of apathy and corpus callosum integrity might reflect slower initiation and longer response times for tasks involving hemispheric metastasis or interregional integration in apathetic ADs. Previous results had demonstrated increased MD in most lobar regions of ADs, including frontal lobes [56], temporal lobes [56], parietal lobes [60], and occipital lobe [60]. DA and RD also increased in addition to FA and MD. However, the relevant research for DA and RD was little. Our findings suggested that DA and RD might be a useful biomarker in identifying HCs and ADs.

After the SVM-RFE method, each kind of DTI metric received the optimal feature dimensions that listed in Table 4. Obviously, the majority of diffusion metrics received optimal effect after dimensionality reduction. In addition, the accuracy, sensitivity and specificity of each kind of diffusion metrics were improved via the SVM-RFE method. Several kinds of multi-metrics were further improved by adjusting the penalty factor C value. Table 4 showed that the FA+MD+RD metric received the best classification accuracy. For further investigated the effect of feature weighting on classification performance, Fig. 3 had been drawn to show the feature weighting distribution of WM tracts by the summation of all combined approaches. The feature weighting of CBF, HG.L and RPIC.L exceed 0.6 while that of ML.R, ALIC.R and PCR.R almost zero. It was also found that the rank of significance of diffusion metrics was FA $>$ DA $>$ MD $>$ RD. In order to visually evaluate classifier output quality, The ROC curve and AUC were depicted from before and after the dimension reduction. Fig. 4 showed that the AUC increased after the SVM-RFE method. The results ulteriorly verified the effectiveness of the dimension reduction.

In this paper, we introduced multi-metrics measures to identify the difference between HCs and ADs based on TBSS method. The corona radiata, corpus callosum and superior longitudinal fasciculus were the WM fiber tracts which mainly suffered the severe damage in ADs. Inter-group classification was completed by SVM-RFE method. Multi-metrics combination would improve the classification performance compared with single diffusion metric. We also depicted the feature weighting distribution of WM tracts for each kind of DTI metric in order to research which WM fiber tract played an important role on classification. In addition, the ROC curve and AUC could evaluate classifier output quality for each kind of diffusion metric.

There were several limitations to this research. First, small sample size would affect the reliability of classification results. Although our research used SVM model to distinguish AD group from HC group, it needed to be further verified on a larger sample to reinforce the current results and ensured that it had strong generalization ability. Deep learning could be combined if necessary. Another limitation was that the MCI group was absent from this study. MCI was known as a transition stage from health status to AD. It was necessary to include MCI in future studies to understand which features develop gradually over the course of the disease evolution and to reveal the degenerative pattern of the pathological mechanism of AD.

The datasets generated and/or analyzed during the current study are available in the ADNI repository, http://adni.loni.usc.edu/.

FZ designed the research study. YZ performed the research, analyzed the data and revised the manuscript. YZ and FZ wrote the manuscript. Both authors contributed to editorial changes in the manuscript. Both authors read and approved the final manuscript. Both authors have participated sufficiently in the work and agreed to be accountable for all aspects of the work.

The dataset we used was the public dataset of ADNI, the ethical approval was not required. In addition, informed written consent was obtained from all participants at every center.

We thank two anonymous reviewers for excellent criticism of the article.

This work was supported by the National Natural Science Foundation of China (Grant No.12202208), the Basic Science (Natural Science) Research Project of Colleges and Universities of Jiangsu Province (Grant No.22KJB310019, Grant No.22KJB130009), the Scientific Research Foundation of high-level personnel of Taizhou University (Grant No.TZXY2021QDJJ001), the Research and Cultivation Project for Young Teachers of Nanjing Audit University (Grant No.2021QNPY015), 2022 Double-Innovation Doctor of Jiangsu Province, “2022 Taizhou Tuo Ju Project” for Young science and Technology Talents, the Project of Excellent Science and Technology Innovation Team of Taizhou University.

The authors declare no conflict of interest.