Vision and convolutional transformers for Alzheimer's disease diagnosis: a systematic review of architectures, multimodal fusion and critical gaps.

The incidence of Alzheimer's disease is rising with the increasing elderly population worldwide. While no cure exists, early diagnosis can significantly slow disease progression. Computer-aided diagnostic systems are becoming critical tools for assisting in the early detection of Alzheimer's disease. In this systematic review, we aim to evaluate recent advancements in computer-aided decision support systems for Alzheimer's disease diagnosis, focusing on data modalities, machine learning methods, and performance metrics. We conducted a systematic review following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. Studies published between 2021 and 2024 were retrieved from PubMed, IEEEXplore and Web of Science, using search terms related to Alzheimer's disease classification, neuroimaging, machine learning, and diagnostic performance. A total of 39 studies met the inclusion criteria, focusing on the use of Magnetic Resonance Imaging, Positron Emission Tomography, and biomarkers for Alzheimer's disease classification using machine learning models. Multimodal approaches, combining Magnetic Resonance Imaging with Positron Emission Tomography and Cognitive assessments, outperformed single-modality studies in diagnostic accuracy reliability. Convolutional Neural Networks were the most commonly used machine learning models, followed by hybrid models and Random Forest. The highest accuracy reported for binary classification was 100%, while multi-class classification achieved up to 99.98%. Techniques like Synthetic Minority Over-sampling Technique and data augmentation were frequently employed to handle data imbalance, improving model generalizability. Our review highlights the advantages of using multimodal data in computer-aided decision support systems for more accurate Alzheimer's disease diagnosis. However, we also identified several limitations, including data imbalance, small sample sizes, and the lack of external validation in most studies. Future research should utilize larger, more diverse datasets, include longitudinal data, and validate models in real-world clinical trials. Additionally, explainability is needed in machine learning models to ensure they are interpretable and reliable in clinical settings. While computer-aided decision support systems show significant promise in improving the early diagnosis of Alzheimer's disease, further work is needed to enhance their robustness, generalizability, and clinical applicability. By addressing these challenges, computer-aided decision support systems could play a key role in the early detection of Alzheimer's disease and potentially reduce health care costs.

Recent advances in retinal imaging pathophysiology have shown a new function for biomarkers in Alzheimer's disease diagnosis and prognosis. The significant improvements in Optical coherence tomography (OCT) retinal imaging have led to significant clinical translation, particularly in Alzheimer's disease detection. This systematic review will provide a comprehensive overview of retinal imaging in clinical applications, with a special focus on biomarker analysis for use in Alzheimer's disease detection. Articles on OCT retinal imaging in Alzheimer's disease diagnosis were identified in PubMed, Google Scholar, IEEE Xplore, and Research Gate databases until March 2021. Those studies using simultaneous retinal imaging acquisition were chosen, while those using sequential techniques were rejected. "Alzheimer's disease" and "Dementia" were searched alone and in combination with "OCT" and "retinal imaging". Approximately 1000 publications were searched, and after deleting duplicate articles, 145 relevant studies focused on the diagnosis of Alzheimer's disease utilizing retinal imaging were chosen for study. OCT has recently been demonstrated to be a valuable technique in clinical practice as according to this survey, 57% of the researchers employed optical coherence tomography, 19% used ocular fundus imaging, 13% used scanning laser ophthalmoscopy, and 11% have used multimodal imaging to diagnose Alzheimer disease. Retinal imaging has become an important diagnostic technique for Alzheimer's disease. Given the scarcity of available literature, it is clear that future prospective trials involving larger and more homogeneous groups are necessary, and the work can be expanded by evaluating its significance utilizing a machine-learning platform rather than simply using statistical methodologies.

BackgroundThe field of Alzheimer disease (AD) has been moving toward earlier detection, personalized assessment of dementia risk, and dementia prevention. In the near future, a gap is expected between the growing demand for Alzheimer-related health care and a shrinking workforce. Responsibility is increasingly assigned to individuals to take an active role in their own brain health management and dementia prevention. Digital tools are thought to offer support regarding these processes.ObjectiveThe aim of this scoping review is to create an overview of digital tools published in scientific literature in the context of AD and dementia aimed at people without an AD or dementia diagnosis as primary end users interacting with these digital tools. Additionally, we aim to gain insight into study sample diversity, the stage of maturity and evaluation of these tools, and recommended future directions.MethodsPubMed, IEEE Xplore, Ovid, and Web of Science were searched in January 2023, using terms related to AD and dementia, (pre-)disease stages, digital tools, and various purposes of digital tools. Two independent reviewers screened the titles and abstracts of 2811 records and subsequently 408 full-text articles, based on inclusion and exclusion criteria. Articles on tools targeting those with an AD or dementia diagnosis were excluded. Data extraction included information on the sample characteristics, the digital tool, stage of maturity and evaluation, and future (research) directions.ResultsWe included 39 articles, which were aimed at primary prevention (14/39, 36%), secondary prevention (11/39, 28%), daily life support (8/39, 21%), self-administered screening (4/39, 10%), or decision-making (2/39, 5%). Variation in the study sample emerged regarding cognitive abilities (healthy: 11/39, 28%; mild cognitive impairment: 12/39, 31%; [subjective] cognitive impairment: 9/39, 23%; “no dementia”: 1/39, 3%; and variation of cognitive abilities: 6/39, 15%). Less variation was found regarding sex (>50% female: 27/39, 69%), education (>50% high education: 13/39, 33%), and age (>50% >60 y: 23/39, 59%). Few articles reported on ethnicity (12/39, 31%) and digital literacy (11/39, 28%). Most tools were in an early evaluation and maturity stage (31/39, 80%), comprising preprototyping (1/35, 3%), prototyping (15/35, 43%), pilot testing (19/35, 54%), efficacy testing (18/40, 45%), usability testing (12/40, 30%), and feasibility testing (10/40, 25%). Future (research) directions comprised the need for further tool development, attention to diversity, and study advancements, such as large-scale longitudinal studies.ConclusionsAlmost 80% of tools as reported on in academic literature are in early development comprising early stages of maturity and evaluation. Studies and evidence gathered for digital tools developed in the context of AD or dementia aimed at people without an AD or dementia diagnosis are thus preliminary and further development, research, and policy are required before these tools can be implemented for assessing, supporting, and preventing cognitive decline.

To the Editor: We would like to address several issues raised by a recent article by Gill et al.1 purporting to review the evidence-based literature regarding the potential value of positron emission tomography (PET) in the diagnosis of dementia. It concludes that there is little evidence to support the integration of PET in the clinical evaluation of patients with suspected or established dementia. Several features of this article and the data reviewed challenge this conclusion. First, the article presents itself as a cost-benefit analysis of PET in the diagnosis of Alzheimer's disease (AD), but no cost-benefit or economic analysis was performed beyond stating the cost of these scans in Ontario, Canada. Such analyses have been conducted using established cost-benefit methods based on imaging and other costs and accounting for such advantages as early introduction of therapy, deferral of nursing home placement, and reduction in use of unnecessary testing. A formally conducted cost-benefit analysis demonstrated that PET is, under the most conservative analysis, at least cost neutral and, under more-realistic conditions, cost advantageous when added to the dementia evaluation.2 Second, the authors stated that they conducted a systematic review of the peer-reviewed literature from Medline from 1975 to 2001, but they did not identify the largest single-institutional study examining the relationship between PET-based and autopsy-based diagnoses, published in 2000.3 In addition, they excluded from their main analysis the largest multicenter study to examine this relationship,4 although its publication fell within the specified time, and they note that this latter article had several advantages over the articles examined. This article demonstrated 94% sensitivity and 73% specificity for PET in the diagnosis of AD, comparable with or better than most clinical diagnostic studies. Third, the authors do not provide criteria for deciding how PET was judged to be beneficial. They conducted no critical assessment or comparisons of sensitivity, specificity, or accuracy measures for clinical diagnoses or PET-based diagnoses. Moreover, they included only papers in their main analysis that had as their primary criterion standard the clinical diagnosis of AD, excluding those papers that used the more definitive standard of autopsy-confirmed diagnosis.3,4 There was consequently no way their analysis could demonstrate any incremental value of PET over clinical diagnosis by the way they constructed and performed their evaluation. Fourth, the authors repeatedly assert that clinical diagnosis of probable AD is straightforward and accurate in up to 90% of cases, thereby seeming to obviate a priori the need for neuroimaging. They identify only one study in support of this claim,5 a paper that the American Academy of Neurology (AAN) recently identified as having Class II quality of evidence. The paper showed that, to achieve a sensitivity of 90% (as occurs with PET), clinical specificity fell to below 40%. Three papers that the AAN rated as having Class I quality of evidence demonstrated a mean accuracy rate of clinical diagnosis of less than 70%. Fifth, Table 3 of the Gill et al.1 article lists the 16 articles or publications included in their review. Of these 16, five were published before the advent of any Food and Drug Administration–approved therapy for AD, and five more were published in 1993 and 1994, when tacrine, a little-used compound, was the only available treatment. Thus, 10 of the 16 articles precede the contemporary era of pharmacotherapeutics in the management of AD. This is important because an early and accurate diagnosis becomes more urgent once therapy is widely available. Sixth, the two Class A or B articles identified by the authors and published in 1996 (the only ones in the current era of neurotherapeutics for AD) were supportive of the use of PET. One study6 found high inter- and intraobserver agreement in PET interpretation of patients with probable AD, possible AD, mild cognitive impairment, and normal controls. Another study7 found that three-dimensional stereotactic surface projections improved sensitivity and specificity. Thus, the two articles published more recently and using more-modern PET scanners support the use of PET in the diagnosis of AD. Seventh, factual errors are present in this article. On four occasions in the article, the authors state that patients diagnosed with probable AD have advanced disease when the diagnosis is easiest and therefore PET adds little to resolve diagnostic challenges. It is untrue that probable AD is necessarily advanced. Probable AD8 refers to patients who meet research criteria for AD. These criteria can be applied as soon as the patient has impairment in memory and in at least one other cognitive domain causing disability. Patients may have Clinical Dementia Rating scale scores as low as 0.5 and still meet criteria for probable AD. Diagnosis of patients at this stage of dementia is a challenge because all dementias necessarily go through mild stages of severity before reaching more-severe and more diagnostically definitive stages. A large multicenter study4 showed the accuracy of PET to be the same at mild and moderate stages of AD. Eighth, the principal challenge in the recognition, diagnosis, and treatment of AD is the lack of recognition of patients by primary care practitioners who are most likely to encounter them in early stages of the disease. Practice reviews show that 97% of patients with early dementia go undiagnosed and as many as 50% of patients with moderate to severe dementia receive no diagnosis.9 Thus, the relatively similar sensitivity and specificities reported by some academic medical centers resulting from the rigorous application of research diagnostic criteria and those of PET are irrelevant to most routine clinical practices. Primary care practitioners fail to address the important issue of cognitive decline in their patients in part because of the absence of an available diagnostic test with which to confirm their opinion. PET should not replace a thorough clinical assessment but can add important and accurate positive evidence to the diagnosis based on traditional evaluations. In summary, the absence of a cost-benefit analysis, the lack of definitions for PET utility, the truncated literature review, the emphasis on out-of-date information, the factual misstatements, and the ignoring of the potential benefit of PET to those who are most likely to benefit from its availability undermine the conclusions of this article. The available literature supports the use of PET in the assessment of dementia, and we recommend that PET be integrated into the diagnostic approach to dementia.10

ObjectivesThere is a lack of studies on the rate and temporal changes of infections in relation to Alzheimer's disease (AD) diagnosis. We studied the infection rate in persons with and without AD yearly 5 years before and after AD diagnosis. DesignRegister-based cohort study. Setting and ParticipantsWe used the Medication Use and Alzheimer's Disease cohort with 70,718 Finnish community dwellers diagnosed with AD between 2005 and 2011 and an equal number of age, sex- and region-of-residence–matched comparison persons. MethodsData on comorbidities, medication use, and hospital days due to infection were retrieved from multiple nationwide registers. The rate of hospitalization and accrued hospital days due to infections were calculated yearly during the follow-up. The accumulation of hospital days was investigated with the negative binomial model. ResultsDuring the follow-up, one-half of persons with AD had inpatient stays due to infections compared with 34% of persons without AD. The infection rate increased substantially 1 to 2 years before AD diagnosis. At AD diagnosis, the rate of inpatient stays and outpatient visits due to infection was higher (15 per 100 person-years) in persons with AD than in comparison persons (9 per 100 person-years), and the accumulation of hospital days in persons with AD was higher a year after the diagnosis (incidence rate ratio, 1.21; 95% CI, 1.11-1.32) due to higher infection rate. The most common infection diagnoses in both groups were pneumonia and genitourinary infections. Conclusions and ImplicationsCompared with matched comparison persons, the higher hospitalization rate due to infections could be caused by systemic inflammation related to AD, infections generally treated in outpatient care, delirium symptoms associated with infections, and caregiver burden. The prevention of infections should be part of the care of cognitive disorders throughout the disease.

Battle against Alzheimer's Disease: The Scope and Potential Value of Magnetic Resonance Imaging Biomarkers

A 3D multi-modal multi-scale end-to-end classifier for Alzheimer's disease diagnosis.

How far is arterial spin labeling MRI from a clinical reality? Insights from arterial spin labeling comparative studies in Alzheimer's disease and other neurological disorders.

BackgroundAn important problem is discovering and understanding the reasons behind Alzheimer's disease (AD) diagnosis from analyzing brain images. We can use computer vision techniques to find the relationship among top‐ranked 2D brain image patches, relevant feature sets, associated brain areas, and the disease diagnosis. A new robust 2D image patch ranking algorithm is created to generate a reliable 2D patch ranking map (PRM) displaying top‐ranked 2D brain image patches for explainable AD diagnosis.MethodAn efficient convolutional neural network (CNN) model consisting of convolutional layers, a new feature selection (FS) layer, and a classifier is built. The new FS algorithm is made to reliably select the top common features from different top feature sets. The new FS‐Grad‐CAM method is developed to generate explainable heatmaps with a smaller number of highlighted areas associated with top‐ranked features. Two new feature matrices and two new heatmap matrices are used to reliably rank patches to better explain the relationship among patches, top‐ranked features, relevant top‐ranked feature maps, and the image classifications.ResultThe AD MRI preprocessed dataset for 4‐class image classification with 6,400 128x28 axial brain images is used for simulations. The associations between brain areas and AD are analyzed by using hybrid information from both relevant publications and ChatGPT. Simulation results show that the top 10 patches (i.e., 3.9% of all 256 patches) are associated with 40.4% of all 57 brain areas associated with AD, 11.4% of all 44 brain areas likely associated with AD, and 0% of all three brain areas not likely associated with AD, as shown in Table 1. The efficient CNN with FS can have higher classification accuracy, smaller model size, and higher explainability than the conventional CNN. Figure 1 shows a PRM with top 10 brain image patches associated with relevant brain areas for AD diagnosis.ConclusionThe new robust 2D image patch ranking algorithm can reliably generate the 2D PRM for interpretable AD diagnosis. A medical doctor may conveniently use the 2D PRM to understand the relationship among top‐ranked 2D image patches, relevant feature sets, important associated brain areas, and AD diagnosis.

Multi-scale multimodal deep learning framework for Alzheimer's disease diagnosis

Trends in health service use and potentially avoidable hospitalizations before Alzheimer's disease diagnosis: A matched, retrospective study of US Medicare beneficiaries

Multi-modal fusion has become an important data analysis technology in Alzheimer's disease (AD) diagnosis, which is committed to effectively extract and utilize complementary information among different modalities. However, most of the existing fusion methods focus on pursuing common feature representation by transformation, and ignore discriminative structural information among samples. In addition, most fusion methods use high-order feature extraction, such as deep neural network, by which it is difficult to identify biomarkers. In this paper, we propose a novel method named deep multi-modal discriminative and interpretability network (DMDIN), which aligns samples in a discriminative common space and provides a new approach to identify significant brain regions (ROIs) in AD diagnosis. Specifically, we reconstruct each modality with a hierarchical representation through multilayer perceptron (MLP), and take advantage of the shared self-expression coefficients constrained by diagonal blocks to embed the structural information of inter-class and the intra-class. Further, the generalized canonical correlation analysis (GCCA) is adopted as a correlation constraint to generate a discriminative common space, in which samples of the same category gather while samples of different categories stay away. Finally, in order to enhance the interpretability of the deep learning model, we utilize knowledge distillation to reproduce coordinated representations and capture influence of brain regions in AD classification. Experiments show that the proposed method performs better than several state-of-the-art methods in AD diagnosis.

Introduction: Alzheimer's disease is one of the most common causes of dementia, which gradually causes cognitive impairment. Diagnosis of Alzheimer's disease is a complicated process performed through several tests and examinations. Design and development of Clinical Decision Support System (CDSS) could be an appropriate approach for eliminating the existing difficulties of diagnosing Alzheimer's disease. Materials and Methods: This study reviews the current problems in the diagnosis of Alzheimer's disease with an approach to the application of CDSS. The study reviewed the articles published from 1990 to 2016. The articles were identified by searching electronic databases such as PubMed, Google Scholar, Science Direct. Considering the relevance of articles with the objectives of the study, 29 papers were selected. According to the performed investigations, various reasons cause difficulty in Alzheimer's diagnosis. Results: The complexity of diagnostic process and the similarity of Alzheimer's disease with other causes of dementia are the most important of them. The results of studies about the application of CDSSs on Alzheimer's disease diagnosis indicated that the implementation of these systems could help to eliminate the existing difficulties in the diagnosis of Alzheimer's disease. Conclusion: Developing CDSSs based on diagnostic guidelines could be regarded as one of the possible approaches towards early and accurate diagnosis of Alzheimer's disease. Applying of computer-interpretable guideline (CIG) models such as GLIF, PROforma, Asbru, and EON can help to design CDSS with the capability of minimizing the burden of diagnostic problems with Alzheimer's disease.

INTRODUCTIONWe aimed to explore primary care physicians’ (PCP) attitudes, perceptions, and barriers toward Alzheimer's disease (AD) diagnosis and incorporating blood biomarker (BBM) tests into the diagnostic workflow.METHODSRemote 60‐min interviews with 20 PCPs were conducted (May 2023). Participants included generalists and geriatricians representing urban, suburban, and rural U.S. practices. Interviews encompassed early AD diagnosis, PCP role, and BBM test implementation.RESULTSMost PCPs view investigating cognitive decline as an important part of their role and are somewhat confident in diagnosing AD. Barriers include the complexity and inefficiency of current diagnostic workflows, lack of effective treatments, and stigma. PCPs consider BBM tests accurate and cost‐effective but have concerns about reimbursement and diagnostic pathway placement.DISCUSSIONPCPs are interested in AD diagnosis and receptive toward BBM testing. Education on BBM test use and AD diagnosis may benefit PCPs in the care of individuals with cognitive decline.HighlightsEarly Alzheimer's disease (AD) diagnosis is crucial for initiating treatmentPrimary care physicians (PCPs) find investigation of cognitive decline importantPCPs consider blood biomarker (BBM) tests accurate and cost‐effectivePCPs seek clarity on reimbursement of BBM tests and their context of useEducation on BBM test interpretation and AD diagnosis may benefit primary care

Objective: To explore the value of multiple visual rating scales based on structural MRI in the diagnosis of Alzheimer's disease (AD) in the Chinese population.Materials and Methods: One hundred patients with AD and 100 age- and gender- matched cognitively normal controls were enrolled in this study. All the participants underwent neuropsychological tests and a structural MRI scan of the brain, among them, 42 AD cases and 47 cognitively normal controls also underwent 3D-T1 weighted sequence used for the analysis of voxel-based morphometry (VBM). The AD cases were divided into mild and moderate–severe groups according to the mini-mental state examination. Each participant was evaluated by two trained radiologists who were blind to the clinical information, according to the six visual rating scales, including for medial temporal lobe atrophy (MTA), posterior atrophy (PA), anterior temporal (AT), orbitofrontal (OF) cortex, anterior cingulate (AC), and fronto-insula (FI). Finally, we estimated the relationship between the visual rating scales and the volume of corresponding brain regions, using correlation analysis, and evaluated the specificity and sensitivity of every single scale and combination of multiple scales in the diagnosis of AD, using a receiver operating characteristic (ROC) curve and establishing a logistic regression model.Results: The optimal cutoff of all six visual rating scales for distinguishing AD cases from normal controls was 1.5. Using automated classification based on all six rating scales, the accuracy for distinguishing AD cases from healthy controls ranged from 0.68 to 0.80 (for mild AD) and 0.77–0.90 (for moderate–severe AD), respectively. A diagnostic prediction model with a combination of MTA and OF results was made as follows: Score = BMTA(score) + BOF(score) −1.58 (age < 65 years); Score = BMTA(score) + BOF(score) −4.09 (age ≥65 years). The model was superior to any single visual rating scale in the diagnosis of mild AD (P < 0.05).Conclusion: Each of the six visual rating scales could be applied to the diagnosis of moderate-severe AD alone in the Chinese population. A prediction model of the combined usage of MTA, OF, and age stratification for the early diagnosis of AD was preliminarily established.