The WoSCC database was chosen as the data source. It is the most frequently used and acceptable database for researchers in a variety of fields.

We searched for publications about MRI in the field of AD on August 26, 2024. The search query string was as follows: TS = (Alzheimer’s disease OR Alzheimer disease OR Alzheimer) AND (magnetic resonance imaging OR MRI OR T1WI OR T2WI OR DWI OR diffusion-weighted imaging OR ESWAN OR enhanced gradient echo T2 star weighted angiography OR SWI OR susceptibility weighted imaging OR MRS OR magnetic resonance spectroscopy OR ASL OR arterial spin labeling OR DCE OR dynamic contrast-enhanced OR PWI OR perfusion weighted imaging OR BOLD-fMRI OR blood oxygenation level-dependent functional magnetic resonance imaging OR DTI OR diffusion tensor imaging OR DKI OR diffusion kurtosis imaging OR IVIM OR intravoxel incoherent motion OR CEST OR chemical exchange saturation transfer OR APT OR amide proton transfer OR MPRAGE OR three-dimensional T1-weighted magnetization-prepared rapid gradient echo OR QSM OR quantitative susceptibility mapping OR rs-fMRI OR resting-state functional MRI OR fMRI OR functional magnetic resonance imaging OR VBM OR voxel-based morphology OR multimodal imaging technique OR volumetric MRI OR structural magnetic resonance imaging) AND (FPY = 2004–2023). The language was limited to English, and the document types were limited to articles and review articles. Information regarding titles, keywords, abstracts, authors, institutions and reference records of the papers was downloaded and saved in plain text format. The study flowchart is shown in Figure 1.

Microsoft Excel 2017 (Microsoft, Redmond, WA, United States) was used to display the global trend of publications and analyze the targeted files export statistical charts and tables of the most productive countries/regions, institutions, authors and journals.

VOSviewer (version 1.6.20) and CiteSpace (version 6.3. R1) are two widely used software programs for constructing and visualizing bibliometric networks (33, 36). CiteSpace was applied to analyze institutions, reference clusters and keyword bursts. VOSviewer was used for the analysis of countries/regions, authors and journals. The node in each map represents a country, institution or reference. The size of the node (country, institution or reference) represents the number of publications. The larger the node, the greater is the number of publications. The links between the nodes represent the strength of collaborations.

A total of 13,659 articles were included in the subsequent analyzes. The number of articles published each year and the cumulative number of articles published are shown in Figure 2, which shows a consistent overall upward trend in the number of annual publications from 2004 to 2022. The fastest annual growth was in 2020, with an increase of 138 articles. The statistical graph is adjusted to fit a linear curve that follows the following equation: the fitted curve index is y = 700.62x – 2230, with a correlation coefficient of 0.9511. In addition, the number of articles published in 2023 decreased from 1,335 in 2022 to 1,276, which may be caused by research bias in that it takes time for articles to reach a certain number of citations after publication.

A total of 108 countries/regions published related articles in this field, 57 of which published no fewer than 10 articles. Table 1 shows the top 10 most productive countries/regions. The United States ranked first with 5,504 articles, accounting for 40.30% of the total number of articles published, followed by China (2,150, 15.74%), England (1,782, 13.05%), and Germany (1,192, 8.73%).

The H-index is a mixed index that is used as a significant indicator of the number and level of academic output of a scientific researcher, country, journal, or institution (37). The country with the highest H-index was the United States, followed by England, Germany and the Netherlands. A cooperative network of countries and their collaborations via VOSviewer is shown in Figure 3.

In addition, 9,233 institutions published articles in this field, of which 1,491 institutions had no fewer than 5 articles. Table 2 summarizes the top 10 institutions and their partnerships. The University of California System published the most, with 1,350 articles, accounting for 9.88%, followed by the University of London (854, 6.25%), Harvard University (738, 5.40%), and Mayo Clinic (625, 4.58%). The cooperative network of institutions and their collaborations via CiteSpace are shown in Figure 4.

A total of 48,736 authors were included in this study, 3,784 of whom published no fewer than 5 articles. Table 3 summarizes the top 10 most productive authors. Half of the authors were from the United States. Among them, Jack CR, Scheltens P, and Petersen RC were the top 3 productive authors, with 413, 240, and 224 articles, respectively.

A total of 1,326 journals published related articles in this field, of which 367 authors published no fewer than 5 articles. The top 15 most prolific journals are listed in Table 4. Among the top 15 journals, one in three were from the United States. In addition, the journal with the highest IF was Brain among the top 15 journals. Journal of Alzheimers Disease [impact factor (IF) 2023: 3.4] published the most (1,175 publications), followed by Neurobiology of Aging (IF 2023: 3.7, 592 publications), and Neuroimage (IF 2023: 4.7, 561 publications). The top 15 journals published 5,460 articles, accounting for 40.12% of the total number of publications. These journals have made great contributions to the development of MRI studies in AD, indicating that more high-quality articles in this field will be published in these journals as a priority in the future.

Stage I focused on the study of morphological changes in the AD brain via sMRI. The main keywords were associated with gray matter loss, the entorhinal cortex, temporal lobe atrophy, hippocampal atrophy, brain atrophy, and white matter lesions.

sMRI is widely used to explore the underlying pathophysiology of AD (38). Brain atrophy detected by sMRI is a valid biomarker of the stage and intensity of AD pathology (9, 10, 39). Structural changes in the brain map accurately upstream to Braak stages of NFT deposition (4, 40, 41) and downstream of neuropsychological deficits (42, 43). Considerable evidence suggests that AD initiates in the entorhinal cortex and hippocampus and spreads thereafter to the rest of the brain. The earliest sites of tau deposition and MRI-based atrophy are located along the perforant (polysynaptic) hippocampal pathway (entorhinal cortex, hippocampus and posterior cingulate cortex), which is consistent with early memory deficits (9, 44, 45). Progressive atrophy in the temporal, parietal and frontal neocortex is closely related to neuronal loss, as well as language, praxic, visuospatial and behavioral impairments (4, 17, 46, 47).

Several different processing methods, such as voxel-based morphometry (VBM) (48), boundary shift integral (BSI) (4, 49), and tensor-based morphometry (TBM) (45, 50), are employed to detect subtle changes and assess atrophy in the brain MR images of the same individual or groupwise comparisons. VBM studies have revealed that regional gray matter atrophy lies mainly in the bilateral mesial temporal lobes, including the hippocampus, amygdala, entorhinal cortex, and posterior cingulum, and extends to the frontal and parietal lobes with the progression of AD (51–53). VBM studies have also revealed white matter abnormalities in patients with AD compared with healthy controls (54–59). Nevertheless, structural neuroimaging studies of white matter volume in AD patients have yielded variable findings owing to heterogeneous subjects or small sample sizes, as well as methodological differences among studies. For example, two studies (56, 57) reported white matter volume reduction only in bilateral temporal structures, whereas other studies (54, 55, 58, 59) reported widespread white matter volume abnormalities beyond the temporal lobe. Li reported that white matter atrophy in AD patients occurred mainly in bilateral structures close to memory formations, such as the hippocampus, amygdala, and entorhinal cortex, through meta-analysis (60).

Stage II focused on the study of biomarkers and fMRI in AD, and the main keywords were biomarkers, transgenic mice, and Pittsburgh compound b.

The major biomarkers of AD that are typically considered for clinical trials and observational studies are cerebrospinal fluid (CSF) Aβ1-42, CSF total-tau (t-tau), fluoro-deoxy-glucose positron emission tomography (FDG–PET), Pittsburgh compound B-PET (PIB–PET), and sMRI. As suggested by Wahlund and Blennow (61, 62), CSF tau, p-tau, and sMRI may reflect the disease stage or intensity of AD, whereas CSF Aβ represents a specific molecular pathway or etiology. These main CSF biomarkers have high diagnostic accuracy for identifying prodromal AD in the mild cognitive impairment (MCI) stage, with a sensitivity and specificity of 85–90% (25, 26). However, variability in measurements between clinical laboratories has hindered the identification of a unified critical value for CSF biomarkers because of differences in analytical procedures for manual ELISA methods between different laboratories, as well as variability in reagent quality and manufacturing procedures resulting in batch-to-batch variations (63). Research based on Alzheimer’s Disease Neuroimaging Initiative (ADNI) data has shown that sMRI is more closely related to cognition than CSF biomarkers are (64, 65).

In earlier stages of AD, such as MCI or preclinical AD, atrophy can be minimal, although metabolic imaging (FDG PET) findings may already be abnormal (66, 67). Although there are many amyloid imaging PET tracers based on ¹¹C and ¹⁸F, the most studied tracer in the field of AD is PIB (68). C-pPIB can be used as a powerful biomarker of rCBF to measure neural activity and improve the diagnostic ability of PET for AD in conjunction with [¹¹C]-aPIB (69). Tiepolt (70) reported that early [¹⁸F]FBB and [¹¹C]PiB PET brain images are analogous to [¹⁸F]FDG PET images of the AD brain and that these tracers could be used as potential biomarkers in place of [¹⁸F]FDG. Furthermore, [¹¹C]-PIB R (1) could serve as a complementary biomarker of neuronal activity and neurodegeneration in addition to the Aβ load given by [¹¹C]-PIB BP(ND) (71). These findings suggest that Aβ deposition measured by PIB is an upstream process, whereas neurodegeneration is a downstream process initiated by Aβ deposition and is more closely related to cognitive decline (10, 72). There are two possible reasons why the keyword “PIB (amyloid PET tracer)” appears in the field of MRI studies in AD. One of which is some studies have evaluated AD using both PET-CT and MRI for pairwise comparison (72). Another potential reason is that partial studies have used PET-MR, a novel imaging technique, to evaluate AD (73).

Genetic data from the ADNI database and from transgenic model experiments have been crucial in advancing the understanding of AD pathophysiology. All of these human genetic data can be obtained from apolipoprotein E (APOE) genotyping, genome-wide association studies (GWAS), and whole-exome and whole-genome sequencing. Researchers have strived to identify new genetic susceptibility loci for AD by integrating genetic and imaging data obtained from the ADNI database. The first GWAS of an ADNI quantitative phenotype was published in 2009 (74). Potkin identified a number of loci potentially associated with hippocampal atrophy through an imaging-genetics approach, and progress has been rapid since then. Furney discovered PICALM, a significant gene associated with entorhinal cortical thickness (75). To date, GWASs have identified more than 20 genetic loci associated with the risk of AD (1, 27, 76). It is well known that APOE ε4 is the major genetic risk factor for late-onset or sporadic AD (77), whereas mutations in amyloid precursor protein (APP), presenilin-1 (PSEN1), and presenilin-2 (PSEN2) can cause early-onset or familial AD (78, 79). It is widely reported that APOE e4 allele is the strongest genetic risk factor of ARIA incidence in anti-Aβ monoclonal antibody-treated AD patients (6, 7, 80). Furthermore, transgenic murine model experiments have revealed the underlying mechanisms as well as treatment of AD in animal models using high-field MRI (22, 28–31, 81). Burgess (28) demonstrated that repeated MR imaging-guided focused ultrasound treatments led to spatial memory improvement in a Tg mouse model of AD, which may be mediated by decreased amyloid pathologic abnormalities and increased neuronal plasticity. Snow (22) demonstrated for the first time that diffusion tensor imaging (DTI) abnormalities were present in the gray matter of an AD mouse model in which both pathological hallmarks are present. Shah (31) identified that hypersynchrony of function connection (FC) may be used as a new non-invasive read-out of early AD and can be recovered by anti-Aβ treatment in TG2576 mice, providing an early therapeutic window before amyloid plaque deposition. Overall, MRI is more widely used in clinical settings for assisting in the early diagnosis and prognosis prediction of AD patients, while preclinical research toward precision medicine and pathological mechanisms in small animal models using MRI is relatively less, because of the high requirements for MR equipment, e.g., ultra-high field MRI (often ≥7.0 T) is generally required.

Furthermore, multimodal MRI is widely used for the early diagnosis of AD. In addition to sMRI and PET-MRI, fMRI has been used to measure the activity of neurons and evaluate the functional relationships between brain regions indirectly and noninvasively. First, increased hippocampal activity has been found to be associated with increased risk for AD (82, 83). Studies have revealed reduced left CA2, 3 and dentate gyrus (CA23DG) activity in cognitively intact APOE ε4 carriers, which may suggest that reduced neural activity in hippocampal subregions may underlie the compensatory increase in extrahippocampal activity in people with a genetic risk for AD prior to the onset of cognitive deficits (84). Second, arterial spin labeling (ASL) is an important MR imaging technique used to assess cerebral blood flow (CBF) quantitatively at the tissue level by magnetically labeling inflowing blood (23). AD patients have consistently shown a reduction in CBF in a posterior parietal region, including the posterior cingulate, angular gyrus, precuneus, and superior parietal gyrus (21, 85–88). Chao reported that abnormal perfusion in the precuneus may predict conversion from MCI to AD (89). Collij demonstrated that the combination of CBF and the entorhinal cortex atrophy score can significantly increase the diagnostic ability of AD, with a prediction probability of 95% (90). The ASL pattern is remarkably similar to the pattern of hypometabolism observed with FDG PET, and both modalities have similar diagnostic performance (23, 88, 91). Third, resting-state fMRI (rs-fMRI) has become one of the most widely used neuroimaging techniques for studying brain function correlated with pathological biomarkers of AD (92, 93). Studies have detected decreased FC in the default mode network (DMN), and the posterior cingulate cortex/pecuneus is the most severely affected brain region (20). DMN is one of the most vulnerable brain networks in AD because of the earliest accumulation of Aβ (19, 20, 94, 95). The early deposition of the pathological protein Aβ in the DMN region may be related to the influence of resting-state whole-brain activity on the lymphatic clearance pathway (96), suggesting that rs-fMRI technology has the potential to be used as an imaging indicator for the early identification of AD and for guiding clinical prevention and treatment. Anti-amyloid monoclonal antibodies have raised concerns about adverse effects, particularly ARIA (5, 6, 8). ARIA includes ARIA-E (parenchymal or sulcal hyperintensities on FLAIR indicative of parenchymal edema or sulcal effusions) and ARIA-H (hypointense regions on gradient recalled-echo/T2* indicative of hemosiderin deposition) (97). Timely detection and monitor of ARIA in clinical practice is crucial.

Stage III focused on the study of AI techniques in the MRI analysis of the AD brain. The research focused on “#0 tau pathology” and “#1 deep learning (DL),” and the main keywords used were subjective cognitive decline, machine learning (ML), convolutional neural network (CNN), and DL.

Recently, increasing evidence has shown that hyperphosphorylated tau protein appears earlier than Aβ does and is more closely related to cognitive impairment, which is deemed to be more sensitive for the early detection of AD (98, 99). Both sMRI and FDG-PET are commonly used for measuring tau-mediated neuronal injury. Emerging MRI modalities such as DTI and rs-fMRI have also shown great potential in capturing changes due to tau pathology (100, 101). Visual assessment (102) or quantification of the hippocampus (103) is the most frequently used biomarker for measuring tau-mediated injury in AD and has been confirmed via several autopsy studies (40). In addition, atrophy is closely associated with tau accumulation in certain brain regions and is related to biomarkers of tau accumulation and brain hypometabolism, such as CSF p-tau/t-tau levels, FDG-PET, or PET imaging with tau ligands (104, 105). At present, most of the studies on tau PET have focused on specific regions of interest, including the occipital lobe, parietal lobe, temporal lobe cortex and parietal lobe cortex, in the AD cohort, which may reflect the correlation between tau protein deposition and cognitive decline (106). Ossenkoppele reported that tau PET was a promising prognostic tool for predicting cognitive decline in preclinical and prodromal stages of AD, that was superior to amyloid PET and MRI (107). Although tau PET imaging has attracted much attention compared with Aβ PET imaging, it still needs to be verified on a large scale to become a reliable tool for AD diagnosis.

In the past 6 years, the keywords have gradually evolved to use AI-powered technologies, particularly ML and DL, combined with MRI to achieve early diagnosis and prognosis of AD (108, 109). DL is an important branch of ML that uses neural networks of simple interconnected units to extract patterns from data to solve complex problems (110). A neural network is the basis of the DL method. Common AI algorithms include logistic regression (LR), support vector machines (SVMs), random forests (RFs), CNNs, and nonconvolutional artificial neural networks (NC-ANNs). ML and DL have a wide range of applications for the early differentiation of normal cognition and AD (110–112). To date, the most widely used image analysis model is CNN (113). CNNs have gained popularity quickly in MRI analysis for AD, first in 2017, with promising performance (114). Ashyam applied DeepBrainNet, a recent DL algorithm with transfer learning, to discriminate neurologic diseases on the basis of MRI brain age and achieved 86% accuracy in differentiating AD patients from healthy controls (HCs) (115). The CNN studies that combined T1WI with FDG-PET data yielded considerably different results (98% vs. 90% accuracy for AD patients vs. HCs) (116). Frizzell demonstrated that CNNs have better performance metrics than the other major algorithm types do in the classification of AD, MCI, and normal aging and the prediction of MCI conversion to AD (111). Sima applied an DL-based assistive software that can automatically detect and quantify ARIA on brain MRI scans, which has the potential to be a clinically important tool to improve safety monitoring and management of patients with AD treated with Aβ-directed monoclonal antibody therapies (117). With the development of AI, more noninvasive examinations, minimally invasive diagnoses, and more safer treatment methods have been developed, suggesting that the application of AI in AD has attracted the attention of many researchers and will be a research hotspot in the next few years.