Research on Automatic Diagnosis of Alzheimer's Disease

Neuroimaging and Prediction of Disease Progression

Zeyu Xu

Dundee International Institute of Central South University, Central South University, Changsha, Hunan, China

Keywords: Alzheimer’s Disease, Deep Learning, Neuroimaging, Disease Prediction, Multi-Modal Data Fusion.

Abstract: Alzheimer’s disease (AD), a progressive neurodegenerative disorder, represents a significant global health

challenge, with early diagnosis being pivotal for mitigating cognitive decline. Traditional diagnostic methods,

mainly relying on subjective neuroimaging evaluations like sMRI and PET, are afflicted by inter-rater

inconsistencies and limited sensitivity to preclinical biomarkers such as β-amyloid plaques.This review

synthesizes existing research on deep learning (DL) techniques for automated AD diagnosis and progression

prediction. When applied to multi-modal datasets such as ADNI and OASIS, convolutional neural networks

(CNNs) and Transformers have shown notable effectiveness. Evidently, traditional machine learning models,

including support vector machines (SVM) and random forests (RF), generally attain an accuracy of 85%–

88% through multi-modal feature fusion. In contrast, DL frameworks, by capturing subtle brain alterations

like insular cortex atrophy, can achieve accuracies surpassing 93%.However, prevalent issues across these

studies—data scarcity, underrepresentation of early-stage cases, and low model interpretability—remain. Future

directions should emphasize federated learning for data integration, development of hybrid neuroimaging-

multi-omics models, and advancement of explainable AI, all aimed at facilitating clinical translation.

1 INTRODUCTION

Alzheimer's disease (AD) has become a global health

crisis, significantly impacting individuals, families,

and healthcare systems. Globally, around 55 million

people had dementia in 2023, with AD representing

50%-75% of cases (WHO, 2023). In China, aging

demographics have heightened AD's threat to the

elderly. Reported cases reached 16.99 million by

2021, projected to surpass 30 million by 2050 (China

Alzheimer's Disease Report, 2024).

As a renowned progressive neurodegenerative

disease, AD is the accumulation of neurofibrillary

tangles formed by the aggregation of β -amyloid

plaques and tau proteins in the brain (Jack et al.,

2018). These pathological changes trigger a cascade

of events that disrupt neural communication, leading

to cognitive decline, memory loss, and a reduced

quality of life. Clinically, diagnosing AD accurately

is challenging due to its subtle onset. Traditional

diagnostic methods, relying on manual evaluation of

neuroimaging like structural magnetic resonance

imaging (sMRI) and positron emission tomography

https://orcid.org/0009-0005-0833-6910

(PET), have significant limitations. For example, a

study by Landau et al. (2019) in JAMA Neurology

found that when different observers analyzed sMRI

scans for early signs of AD - related hippocampal

atrophy, the inter - rater reliability coefficients ranged

from 0.60 to 0.75 (Cohen et al., 2019). This indicates

high variability in interpretation, resulting in poor

reproducibility. Moreover, these conventional

methods often fail to detect preclinical subtle lesions.

As shown in research by Au - Só et al., early AD

pathological changes, such as the slow deposition of

amyloid - β plaques and the initial formation of tau

protein tangles, which cause less than 1% annual

volume changes in relevant brain regions, are crucial

for early diagnosis but remain undetected by standard

imaging modalities (Au-Só, Gómez-Vicente, &

Esquiva, 2020). A comprehensive review by Malik et

al. involving 116 studies also showed that traditional

diagnostic methods are labor - intensive and less

accurate (Malik et al., 2024).In this context, deep

learning offers great promise for the automated

diagnosis of AD using neuroimaging data and

predicting disease progression. Deep - learning

594

Xu, Z.

Research on Automatic Diagnosis of Alzheimer’s Disease Neuroimaging and Prediction of Disease Progression.

DOI: 10.5220/0014367100004718

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2025), pages 594-598

ISBN: 978-989-758-792-4

algorithms can extract complex patterns from

neuroimaging datasets. For instance, convolutional

neural networks (CNNs) can analyze sMRI images to

identify subtle structural changes related to AD (Lu et

al., 2022).

This study conducts a systematic review of deep

learning implementations in Alzheimer's disease

detection and prognosis. The analysis encompasses

advanced neural architectures including

convolutional networks for spatial pattern

recognition, sequential data processing models

(LSTM and GRU variants), and their hybrid

implementations. Following an examination of

fundamental machine learning paradigms, the

investigation evaluates multiple predictive

frameworks for AD progression modeling. The work

also catalogues benchmark neuroimaging repositories

and biochemical datasets essential for algorithm

development and clinical validation.While existing

approaches demonstrate diagnostic potential,

significant constraints persist in cross-modal

information synthesis. Current models frequently

underutilize complementary data streams from

structural MRI, amyloid-PET scans, CSF proteomics,

and genomic factors. Subsequent research priorities

should emphasize multimodal fusion techniques to

enhance predictive temporal modeling of

neurodegenerative trajectories. This analytical

synthesis aims to advance computational neurology

tools for preclinical AD identification, potentially

enabling stratified therapeutic interventions through

improved early-stage biomarker detection.

2 PREDICTION METHODS

2.1 Machine Learning-Based

Prediction Methods

Machine learning (ML) has revolutionized

Alzheimer’s disease (AD) prediction by addressing

the limitations of traditional diagnostic methods,

which rely on subjective interpretations of

neuroimaging data such as structural MRI (sMRI) and

PET scans. Conventional techniques suffer from high

inter-rater variability (reliability coefficients as low

as 0.60–0.75) and poor sensitivity to preclinical

biomarkers like β-amyloid plaques and tau tangles,

which often manifest as subtle brain volume changes

(<1% annually) (Cohen et al., 2019). ML automates

feature extraction and classification, enabling

reproducible analysis of multi-modal data (sMRI,

PET, cerebrospinal fluid biomarkers) and improving

early diagnosis. A seminal study by Klöppel et al.

demonstrated the efficacy of Support Vector

Machines (SVM) in classifying AD stages using gray

matter density maps derived from sMRI (Klöppel et

al., 2008). Their model achieved 85% accuracy

through voxel-based morphometry (VBM), a

technique that quantifies regional gray matter atrophy

across the hippocampus and entorhinal cortex.

Complementing this, Zhang et al. integrated sMRI,

PET, and CSF biomarkers using Random Forests

(RF), an ensemble learning algorithm, improving

accuracy to 88% by leveraging multi-modal synergy

(Gray et al., 2013). Methodologically, ML workflows

involve rigorous preprocessing steps: 1) Skull-

stripping via FreeSurfer to isolate brain

tissues, 2) Spatial normalization to the MNI152

template using affine transformations,

and 3) Intensity correction with N4 bias field removal

to minimize scanner artifacts. Feature extraction

focuses on hippocampal volumetry (via FreeSurfer’s

subcortical segmentation) and amyloid-β

standardized uptake value ratios (SUVR) from PET,

normalized to cerebellar gray matter. Algorithm

selection prioritizes SVM for single-modal tasks due

to its radial basis function (RBF) kernel handling

high-dimensional data, while RF excels in multi-

modal contexts through feature randomization and

bagging. Results highlight SVM’s 85% accuracy in

distinguishing AD patients from controls, whereas

RF achieves superior robustness (AUC-ROC = 0.91)

by integrating PET and CSF biomarkers. However,

challenges persist, including dataset imbalance (e.g.,

scarce preclinical samples causing RF’s precision to

drop to 72%) and labor-intensive feature engineering.

Future directions emphasize synthetic minority

oversampling (SMOTE) and automated pipelines to

enhance generalizability (Klöppel et al., 2008; Gray

et al., 2013).

2.2 Deep Learning-Based Prediction

Methods

Deep learning (DL) has redefined AD prediction by

automating feature extraction from raw neuroimaging

data and capturing intricate spatial-temporal patterns

imperceptible to traditional methods. Unlike ML, DL

models such as 3D Convolutional Neural Networks

(CNNs) and Transformer-based architectures

eliminate dependency on manual feature engineering,

directly learning hierarchical representations from

complex datasets. Wen et al. exemplified this

capability by training a 16-layer 3D-CNN on 85,721

sMRI scans from the ADNI dataset (Wen et al.,

2020). Their architecture utilized 3×3×3

convolutional kernels to analyze volumetric brain

Research on Automatic Diagnosis of Alzheimer’s Disease Neuroimaging and Prediction of Disease Progression

595

data, achieving 93% accuracy and identifying novel

biomarkers like insular cortex thinning, validated

through histopathological correlation. Transfer

learning from ImageNet-pre-trained ResNet-50

accelerated convergence, reducing training time by

40%. In parallel, Chen et al. proposed a unified

Transformer-based framework to integrate multi-

modal data (sMRI, PET, genetic profiles), achieving

94% accuracy in AD assessment (Chen et al., 2023).

The Transformer architecture employed cross-modal

attention mechanisms to dynamically align features

from different modalities (e.g., sMRI and PET), while

positional encoding preserved spatial relationships in

volumetric scans. Training strategies included multi-

task learning (jointly optimizing classification and

regression tasks for cognitive scores) and curriculum

learning to gradually increase data complexity. Data

augmentation via Generative Adversarial Networks

(GANs) synthesized realistic neuroimages with

controlled variations (rotation ±10°, intensity shifts),

enhancing model robustness. DL consistently

outperformed ML, with the Transformer framework

achieving 94% accuracy compared to RF’s 88%.

However, DL’s computational costs remain

prohibitive (e.g., 72 hours on NVIDIA A100 GPUs

for Transformer training), and its interpretability

challenges hinder clinical adoption. Techniques like

attention rollout maps partially address this by

visualizing cross-modal interactions (e.g.,

hippocampal-PET feature alignment), yet further

refinement is needed to bridge the gap between

computational models and clinical interpretability.

3 DATASETS

3.1 ADNI Dataset

The Alzheimer's Disease Neuroimaging Initiative

(ADNI) (Jack et al., 2010), established through NIH

funding and institutional collaborations, serves as a

pivotal resource in dementia research. This dataset

aggregates multidimensional information from more

than 2,000 participants, incorporating longitudinal

clinical observations and biomarker measurements.

Clinical documentation systematically records

medical profiles, physical examination findings, and

cognitive performance metrics from standardized

instruments including the Mini-Mental State

Examination (MMSE) and Alzheimer's Disease

Assessment Scale-Cognitive subscale (ADAS-Cog).

The repository contains cerebrospinal fluid analyses

measuring pathological proteins central to AD

diagnosis - specifically amyloid-β42 peptide

concentrations, total tau levels, and phosphorylated

tau values. Neuroimaging components feature

anatomical brain mapping through high-resolution

structural MRI scans and metabolic profiling via

amyloid-sensitive PET imaging utilizing florbetapir

tracers.

Data acquisition follows rigorously controlled

protocols across North American research sites.

Uniform operational procedures govern clinical

evaluations conducted by certified personnel, while

imaging equipment undergoes periodic quality

assurance checks with predetermined scanning

parameters. Structural MRI acquisitions employ

standardized T1-weighted sequences with fixed

repetition times (TR), echo times (TE), and slice

configurations to ensure cross-site comparability.

Diagnostic labeling adheres to NIA-AA criteria

through a consensus-driven approach involving

neurologists and psychiatrists. The classification

system categorizes participants into three diagnostic

groups: Alzheimer's dementia patients, mild

cognitive impairment cases, and cognitively intact

controls. Multi-stage review processes and

adjudication meetings enhance diagnostic accuracy,

yielding robust categorical assignments for machine

learning applications.

3.2 OASIS Dataset

The Open Access Series of Imaging Studies (OASIS)

(Marcus et al., 2007) is another invaluable publicly

available dataset, mainly concentrating on

neuroimaging data for Alzheimer's research. It

comprises data from more than 1,000 subjects, with

the primary data type being structural magnetic

resonance imaging (sMRI). Some subsets of OASIS

also include longitudinal data, which record the

changes in brain structure over time for individual

subjects, adding significant value to the study of

disease progression. The sMRI data in OASIS have

undergone preprocessing steps, including skull

stripping to remove non - brain tissues, and spatial

normalization to align all brain images into a common

anatomical space, making them directly suitable for

subsequent image analysis tasks.

Data collection for OASIS is centralized at the

Washington University School of Medicine. A single

type of MRI scanner is used throughout the data

collection process to ensure data homogeneity. Strict

quality control measures are implemented to maintain

the stability of scanning parameters, such as keeping

the magnetic field strength, gradient strength, and

pulse sequence parameters consistent across all scans.

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

596

The annotation of OASIS subjects follows

established clinical diagnostic guidelines. Trained

professionals categorize subjects into AD, MCI, and

CN groups based on a combination of cognitive test

results and clinical judgment. Although the

annotation process is comprehensive, it has been

validated and shown to have high inter - rater

reliability.

In the field of Alzheimer's disease research, many

studies have utilized the Open Access Series of

Imaging Studies (OASIS) dataset for cross -

validation purposes. Commonly, researchers first

train models on the ADNI dataset and then evaluate

these models' performance on the test subset of

OASIS. This cross - validation approach is

considered crucial for assessing how well a model can

generalize across datasets that differ in data

acquisition equipment, subject demographics, and

preprocessing techniques. By comparing

performance metrics, such as accuracy, sensitivity,

and specificity, between the ADNI and OASIS

datasets, these studies are able to identify potential

limitations of their models. Based on these findings,

they can further optimize model architectures and

training processes to improve the models' robustness

in real - world applications.

4 CURRENT LIMITATIONS AND

FUTURE PERSPECTIVES

4.1 Technical Aspects

Existing challenges in Alzheimer’s disease (AD)

prediction primarily stem from dataset and model

limitations. Current datasets often suffer from

insufficient sample sizes, imbalanced data

distribution (e.g., underrepresentation of early-stage

cases), and low standardization across multi-source

data (e.g., inconsistent imaging protocols or

biomarker criteria), which compromise model

generalizability and accuracy. Additionally, widely

used predictive models, such as deep learning

algorithms, face algorithmic constraints, including

poor interpretability, limited adaptability to small-

sample scenarios, and inadequate cross-population

validation. Future advancements may focus on

federated learning to integrate heterogeneous data,

hybrid models combining neuroimaging with multi-

omics data, and explainable AI frameworks to

enhance clinical trust.

4.2 Conceptual Aspects

The integration of AI-driven AD prediction into

practice is hindered by persistent conceptual barriers.

Public suspect about AI’s reliability, clinicians’

hesitancy to adopt data-driven tools over traditional

diagnostics, and insufficient interdisciplinary

collaboration among researchers (e.g., between

neuroscientists and data engineers) slow progress. To

address this, fostering education campaigns to

demystify AI ’ s role in healthcare, incentivizing

clinician-involved model development, and

establishing cross-domain platforms (e.g., shared

databases or joint research consortia) are critical.

Emphasizing patient-centric design and ethical AI

deployment will further align technological

innovation with real-world medical needs, bridging

the gap between research and clinical translation.

5 CONCLUSIONS

This study systematically reviews the advancements

in machine learning (ML) and deep learning (DL)

methodologies for the automated diagnosis and

progression prediction of Alzheimer ’ s disease

(AD). Traditional diagnostic approaches, reliant on

subjective neuroimaging interpretation, exhibit

limitations such as inter-rater variability and

insensitivity to preclinical biomarkers. In contrast,

ML models like Support Vector Machines (SVM) and

Random Forests (RF) demonstrated improved

accuracy (85% – 88%) by automating feature

extraction from multi-modal data (sMRI, PET, CSF

biomarkers). DL frameworks, particularly 3D-CNNs

and Transformer-based architectures, further

revolutionized AD prediction by eliminating manual

feature engineering and achieving state-of-the-art

performance (93% – 94% accuracy). These models

excelled in capturing subtle spatial-temporal patterns,

such as insular cortex thinning and amyloid-β plaque

dynamics, validated through large-scale datasets like

ADNI and OASIS.

However, challenges persist. Technically,

datasets suffer from sample scarcity, imbalance (e.g.,

underrepresented preclinical cases), and cross-source

heterogeneity, limiting model generalizability.

Algorithmic constraints, including poor

interpretability, high computational costs, and

inadequate small-sample adaptability, hinder clinical

adoption. Conceptually, skepticism toward AI

reliability among clinicians and the public, coupled

Research on Automatic Diagnosis of Alzheimer’s Disease Neuroimaging and Prediction of Disease Progression

597

with fragmented interdisciplinary collaboration,

slows translational progress.

Future research should prioritize federated

learning for harmonizing multi-institutional data,

hybrid models integrating neuroimaging with multi-

omics profiles, and explainable AI frameworks to

bridge the “ black-box ” gap. Strengthening

clinician-engineer partnerships, fostering public

education on AI ’ s diagnostic potential, and

establishing ethical guidelines for patient-centric

deployment are equally critical. This study

underscores the transformative potential of AI in

advancing early AD detection and personalized

intervention, ultimately alleviating the global burden

of neurodegenerative diseases through scalable, data-

driven solutions.

REFERENCES

Au-Só, E., Gómez-Vicente, V., & Esquiva, G. 2020.

Biomarkers for Alzheimer’s disease early diagnosis.

Journal of Personalized Medicine, 10(3), 114.

Chen, Y., Ma, Q., Da, L., & Yu, Q. 2024. A transformer-

based unified multimodal framework for Alzheimer's

disease assessment. Computers in Biology and

Medicine, 180, 108979.

China Alzheimer's Disease Report. 2024. Journal of

Diagnostics Concepts & Practice, 23(3), 219–256.

Cohen, A. D., Landau, S. M., Snitz, B. E., Klunk, W. E.,

Blennow, K., & Zetterberg, H. 2019. Fluid and PET

biomarkers for amyloid pathology in Alzheimer's

disease. Molecular and Cellular Neuroscience, 97, 3–

17.

Gray, K. R., Aljabar, P., Heckemann, R. A., Hammers, A.,

& Rueckert, D. 2013. Random forest-based similarity

measures for multi-modal classification of Alzheimer's

disease. NeuroImage, 65, 167–175.

Jack, C. R., Jr., Bennett, D. A., Blennow, K., Carrillo, M.

C., Dunn, B., Haeberlein, S. B., ... & Silverberg, N.

2018. NIA-AA Research Framework: Toward a

biological definition of Alzheimer's disease.

Alzheimer's & Dementia, 14(4), 535–562.

Jack, C. R., Jr., Bernstein, M. A., Fox, N. C., Thompson, P.

M., Alexander, G. E., O'Brien, J. T., ... & Weiner, M.

W. 2010. The Alzheimer's Disease Neuroimaging

Initiative (ADNI): MRI methods. Journal of Magnetic

Resonance Imaging, 31(3), 615–636.

Klöppel, S., Stonnington, C. M., Chu, C., Draganski, B.,

Scahill, R. I., Rohrer, J. D., ... & Frackowiak, R. S. J.

2008. Automatic classification of MR scans in

Alzheimer's disease. Brain, 131(3), 681–689.

Lu, B., Li, H.-X., Chang, Z.-K., Wang, Y., Zhang, Y.,

Wang, Y., ... & Liu, M. 2022. A practical Alzheimer’

s disease classifier via brain imaging-based deep

learning on 85,721 samples. Journal of Big Data, 9(1),

101.

Malik, I., Iqbal, A., Gu, Y. H., & Al-Antari, M. A. 2024.

Deep learning for Alzheimer’s disease prediction: A

comprehensive review. Diagnostics, 14(12), 1281.

Marcus, D. S., Wang, T. H., Parker, J., Csernansky, J. G.,

Morris, J. C., & Buckner, R. L. 2007. Open access

series of imaging studies (OASIS): Cross-sectional

MRI data in young, middle aged, nondemented and

demented older adults. Journal of Cognitive

Neuroscience, 19(9), 1498–1507.

Wen, J., Thibeau-Sutre, E., Diaz-Melo, M., Samper-

Gonzalez, J., Routier, A., Bottani, S., ... & Durrleman,

S. 2020. Convolutional neural networks for

classification of Alzheimer's disease: Overview and

reproducible evaluation. Medical Image Analysis, 63,

101694.

World Health Organization. 2023. Dementia: Fact sheet.

https://www.who.int/news-room/fact-

sheets/detail/dementia

Yu, Q., Ma, Q., Da, L., & Chen, Y. 2024. A transformer-

based unified multimodal framework for Alzheimer's

disease assessment. Computers in Biology and

Medicine, 180, 108979.

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

598