A Reproducibility-Focused Comparison of CNNs and Vision Transformers for Alzheimer’s Disease Classification from Brain MRI

Authors

  • Sahil Sharma Sahil Sharma Guru Kashi University image/svg+xml Author
  • Dr.Rajinder Kumar Author

DOI:

https://doi.org/10.5281/zenodo.24390098

Keywords:

Alzheimer’s disease, brain MRI, convolutional neural networks, vision transformers, ADNI, participant-level validation, calibration, reproducibility, medical image analysis.

Abstract

Alzheimer’s disease is a progressive neurodegenerative disorder and a major contributor to the global burden of dementia. Structural T1-weighted MRI can capture anatomical correlates associated with disease progression, including medial temporal and hippocampal atrophy, cortical thinning, andventricular enlargement.These patterns are not individually diagnostic, but they provide quantitative information that can complement clinical assessment and other biomarkers [1], [8].

Deep learning has become a common approach to image-based classification. CNNs learn local patterns through spatially shared filters and hierarchical feature extraction. Residual connections support the optimization of deep networks[3],dense connectivity encourages feature re use[4], and compound scaling offers an efficient accuracy–compute trade-off [5]. Vision Transformers divide an image into patches and use self attention to model interactions across the representation[6].Swin Transformer introduces hierarchical representations and shifted-window attention to reduce the cost of global attention while preserving cross-window information [7].

Comparisons between these model families are meaningful only when the experimental conditions are aligned.Ifonearchitecturereceivesalargercohort,different preprocessing, more tuning trials, or a more favorable split, measured differences cannot be attributed confidently to architecture. Neuroimaging also presents a specific leakage risk: multiple scans or slices from the same person can be highlycorrelated.Splittingimagesratherthanparticipantscan allow subject-specific information to appear in both training and test data and inflate apparent generalization.

This study organizes a controlled comparison around a common cohort, MRI preparation pipeline, participant-level split,tuningbudget,andmetricset.Itsintendedcontributions are: (1) a transparent five-model comparison spanning CNN and transformer families; (2) participant-level uncertainty estimation;(3)complementaryassessmentofdiscrimination, calibration,andcomputationalefficiency;and(4)areporting structurethatmakesunresolvedreproducibilityrequirements explicit. The manuscript does not claim that the reported scores establish clinical readiness.

Additional Files

Published

2026-10-12

Issue

Section

Computing and Information Technology

How to Cite

A Reproducibility-Focused Comparison of CNNs and Vision Transformers for Alzheimer’s Disease Classification from Brain MRI. (2026). GKU Journal of Multidisciplinary Research, 2(IV), 63-70. https://doi.org/10.5281/zenodo.24390098