Related Works: A Synthesis of the Literature on Machine Learning–Based MRI Analysis for the Early Diagnosis of Alzheimer's Disease
Literature Review – Aegis-AD: Automated Early Grading and Identification System for Alzheimer's Disease
The computational analysis of structural Magnetic Resonance Imaging (sMRI) for the early identification of Alzheimer's Disease (AD) has matured into a richly multidisciplinary field, situated at the confluence of clinical neurology, computer vision, statistical learning, and medical informatics. Over the last fifteen years, the literature has progressively shifted from purely descriptive radiological observation toward fully automated, data-driven decision-support pipelines capable of stratifying cognitively normal subjects (HC), individuals affected by Mild Cognitive Impairment (MCI), and patients with established AD. This review consolidates fifteen seminal contributions that, taken together, articulate the methodological backbone upon which the proposed Aegis-AD system is constructed. The discussion is organised thematically rather than chronologically, in order to expose the conceptual lineage that connects neuroanatomical biomarker discovery, classical statistical learning, deep representation learning, and modern multimodal ensemble architectures.
Any computational framework intended to detect AD from neuroimaging data must first be grounded in a precise understanding of the macroscopic anatomical signatures of the disease. The foundational work of Frisoni et al. (2010) remains canonical in this regard, framing structural MRI as the de facto in-vivo proxy for neurodegeneration and consolidating Medial Temporal Atrophy (MTA) as the gold-standard radiological marker of disease progression. Critically, Frisoni and colleagues argue that MTA is not a primary lesion but rather a downstream consequence of the underlying amyloid-tau cascade, an interpretation that has profound implications for any learning algorithm: the model is, in effect, quantifying the macroscopic residue of synaptic loss rather than the molecular pathology itself.
This pathophysiological framing is reinforced by the broader literature on structural MRI in the Alzheimer cascade, which establishes that ventricular enlargement and cortical thinning constitute the ex-vacuo signature of dendritic and neuronal death. Henry-Feugeas (2007) further refines this picture by drawing a methodologically essential distinction between the diffuse, symmetrical atrophy associated with normal senescence and the asymmetric, logarithmically severe medial-temporal involvement that characterises the genuine Alzheimer syndrome — most prominently at the level of the amygdala and the entorhinal cortex. This distinction is not merely descriptive: it directly anticipates one of the principal sources of false positives in elderly cohorts, namely the contamination of pathological patterns by benign age-related atrophy.
Building upon these clinical premises, Desikan et al. (2009) delivered one of the earliest fully automated, atlas-driven quantifications of cortical sub-regions, leveraging the FreeSurfer pipeline to demonstrate that the entorhinal cortex — together with the hippocampus — constitutes the most sensitive and earliest-detectable autonomous biomarker of the MCI-to-AD conversion trajectory. Their logistic regression analyses, conducted over carefully curated a priori regions of interest (ROIs), provided robust statistical evidence that an automated morphometric pipeline can rival expert manual delineation, thereby legitimising the entire family of region-based machine-learning approaches that would dominate the subsequent decade.
Whilst structural T1-weighted imaging remains the workhorse modality, Chandra et al. (2019) offer a critical evaluation of complementary techniques — Diffusion Tensor Imaging (DTI), functional MRI (fMRI), and Arterial Spin Labelling (ASL) — and reach a pragmatically important conclusion: although DTI and resting-state fMRI reveal early disruption of the Default Mode Network (DMN), classical volumetric structural MRI remains the only modality endorsed for routine clinical deployment. This finding is decisive for any system aspiring to real-world applicability, since it justifies the deliberate confinement of the analysis to widely available T1-weighted sequences. Complementing the methodological taxonomy, the comprehensive review by Mirzaei et al. (2016) systematises the canonical processing pipeline — acquisition → pre-processing → ROI extraction → classification — and consolidates the empirical primacy of hippocampal volumetry and white-/grey-matter (WM/GM) ratio analysis as the most reproducible feature families, whilst identifying the Support Vector Machine (SVM) as the most enduring baseline classifier in the field.
Taken together, these six contributions establish the clinical and anatomical scaffolding that any automated AD-detection system must respect: the model's attention should converge on the medial temporal lobe, the entorhinal cortex, the hippocampus, and the amygdala, whilst remaining robust to the confounding effect of physiological aging.
Having delineated the anatomical targets, the literature turns to the question of which statistical learners best translate volumetric and morphometric descriptors into reliable diagnostic outputs. The work of Battineni et al. (2020) offers a particularly instructive comparative study in this respect. Operating on a longitudinal cohort of 150 elderly subjects (373 MRI acquisitions) from Washington University, the authors benchmark four canonical models — Naïve Bayes (NB), Artificial Neural Networks (ANN), k-Nearest Neighbours (KNN), and Support Vector Machines (SVM) — under both manually curated and automatically selected feature regimes. Their results are methodologically revealing: whereas the ANN dominated under manual feature engineering (ROC = 0.812), the application of wrapper-based automatic feature selection allowed the comparatively simple Naïve Bayes classifier to achieve a markedly superior ROC of 0.942. Most strikingly, the fusion of all base learners into a hybrid ensemble elevated the area under the ROC curve to 0.991, a result that empirically vindicates the long-standing principle that diverse weak learners, properly aggregated, can outperform any single sophisticated model — particularly in the small-sample regimes characteristic of medical imaging.
A conceptually adjacent but methodologically distinct contribution is offered by Moradi et al. (2015), who address the more clinically demanding task of forecasting MCI-to-AD conversion within a one- to three-year horizon. Recognising that labelled longitudinal data are intrinsically scarce, the authors introduce a semi-supervised learning strategy that exploits unlabelled MRI scans to better delineate the natural decision manifold of the disease. By integrating Voxel-Based Morphometry (VBM) descriptors of cortical density and thickness with cognitive and demographic covariates, their framework attains a degree of temporal stability that purely supervised approaches struggle to match. The contribution is particularly relevant for any system intending to operate on heterogeneous, partially annotated clinical archives.
A further refinement of the classical paradigm is presented by Castellazzi et al. (2020), who confront one of the most clinically vexing problems in dementia diagnostics: the differential diagnosis between Alzheimer's Disease and Vascular Dementia (VaD). Their pipeline selectively feeds tree-based ensembles — most notably Random Forests — with morphometric descriptors enriched by White Matter Hyperintensity (WMH) quantifications and focal cortical atrophy patterns. The principal finding is that an appropriately featured ensemble can disentangle the vascular and neurodegenerative phenotypes by attending to white-matter lesion topology rather than to hippocampal volume alone, a result that explicitly argues against any monolithic, hippocampus-centric model.
Collectively, this triad of works substantiates three robust methodological convictions: (i) wrapper-based feature selection materially improves classical learners; (ii) semi-supervised regimes are not optional in clinical contexts where labels are partial and longitudinal; and (iii) the diagnostic landscape contains differential-diagnosis sub-problems that demand attention to features beyond the canonical medial-temporal signature.
Whereas the classical literature relies on hand-crafted morphometric descriptors, the deep-learning era has progressively automated the very act of feature extraction. The hybrid framework proposed by AlSaeed and Omar (2022) is paradigmatic of this transition. Rather than training a deep network end-to-end — an approach that is notoriously data-hungry and prone to overfitting on medical cohorts — the authors employ a pre-trained ResNet-50 exclusively as a deep feature extractor, redirecting the activations of its penultimate fully-connected layer to three downstream classical classifiers: Softmax, SVM, and Random Forest. Evaluated on the ADNI and MIRIAD cohorts, the architecture attained accuracies in the 85.7%–99% range, with SVM and Softmax exhibiting the most stable behaviour over the ResNet representation. The methodological lesson is unambiguous: the convolutional backbone should be regarded as a learned, task-agnostic morphometric encoder, while the classification responsibility is best delegated to a well-regularised statistical learner.
This view receives independent corroboration from Salvatore and Nanni et al. (2020), who explicitly stage the contemporary methodological duel between transfer-learning-based deep extractors (typically pre-trained on ImageNet) and conventional radiomic descriptors (such as Grey-Level Co-occurrence Matrix features and Discrete Wavelet Transform coefficients). When both feature families are passed through identical SVM classification heads, the deep representations enjoy a clear practical advantage — they obviate manual filter design — yet the classical descriptors retain a non-trivial competitive edge in interpretability and small-sample stability. The convergence of these two studies justifies the now-prevalent deep-to-classic hybrid paradigm, in which deep networks supply the representation and traditional learners provide the decision boundary.
It is worth noting, in contrast to the purely deep approaches, that neither study advocates for the wholesale replacement of classical pipelines; rather, both argue for an architectural symbiosis that preserves interpretability whilst harnessing the representational power of convolutional backbones — a trade-off of particular salience in regulated clinical contexts.
A persistent obstacle in MRI-based learning is the sheer dimensionality of the input domain: a single T1-weighted volume comprises millions of voxels, and unconstrained voxel-wise modelling rapidly succumbs to the curse of dimensionality. Salvatore et al. (2015) offered an early and methodologically elegant solution by coupling Principal Component Analysis (PCA) with a Fisher Discriminant Ratio (FDR) filter prior to SVM classification. Operating on 509 ADNI subjects, their pipeline distilled whole-brain voxel maps — together with grey- and white-matter probability maps — into a compact, statistically discriminative subspace. Beyond the engineering benefit of tractable memory footprints, the resulting "neuroanatomical signature" implicated not only the hippocampus but also the entorhinal cortex, basal ganglia, precuneus, and — somewhat counter-intuitively — the cerebellum, thereby refining the spatially restrictive view inherited from earlier hippocampus-centric work.
A conceptually adjacent contribution is offered by Zhang et al. (2015), who extend the spirit of PCA-based decomposition through the Eigenbrain formalism — a direct three-dimensional generalisation of the eigenfaces paradigm well established in computer vision. By coupling Eigenbrain components with Particle Swarm Optimization (PSO) and kernel SVM classifiers, the authors achieve both high spatial fidelity (mapping atrophy across the temporal lobe, hippocampus, and lateral ventricles) and substantial computational efficiency. The Eigenbrain construct is particularly attractive for didactic and reproducibility purposes, as it grounds the dimensionality-reduction step in an interpretable linear-algebraic formalism.
The aspiration toward generalisable, multi-cohort classifiers reaches its most ambitious articulation in the work of Diogo et al. (2022), who explicitly target a multi-diagnostic and cross-protocol setting. Drawing on both ADNI and OASIS cohorts and combining morphometric descriptors with Graph Theory (GT) metrics, the authors construct a voting ensemble that aggregates linear SVMs, decision trees, and random forests. The framework achieved a balanced accuracy of 90.6% in the HC-versus-AD task and, crucially, demonstrated that models trained on MPRAGE sequences transfer effectively to IR-SPGR acquisitions and vice versa — a non-trivial validation of cross-protocol robustness. Two ancillary findings deserve emphasis: hippocampal volumes accounted for between 25% and 45% of the decision contribution, reaffirming the anatomical primacy of the medial temporal lobe; and graph-theoretical descriptors did not yield statistically meaningful gains, a sober reminder that methodological complexity is no substitute for biomarker validity.
Finally, the multimodal extension proposed by Battineni et al. (2021) complements the imaging-centric literature by demonstrating that the concatenation of MRI-derived feature vectors with structured demographic and clinical-history variables substantially reduces false-positive rates, particularly in the diagnostically ambiguous mild impairment stage. From an engineering perspective, this argues for a late-fusion architecture in which an imaging pipeline and a tabular-data pipeline produce complementary embeddings that are concatenated upstream of the final classifier — a design pattern that aligns naturally with the modular philosophy of Aegis-AD.
The fifteen contributions reviewed above describe a literature that has converged, with remarkable consistency, on a set of methodological and clinical principles. First, the medial temporal lobe, and in particular the hippocampus and entorhinal cortex, constitute the anatomical loci upon which any credible classifier must demonstrably focus. Second, classical T1-weighted structural MRI — despite the emergence of DTI, fMRI, and ASL — remains the only modality with the regulatory and logistical maturity required for routine clinical deployment. Third, the contemporary state of the art has not crowned a single dominant algorithm; rather, it has stabilised around a hybrid architectural pattern in which deep convolutional backbones (e.g., ResNet-50) operate as task-agnostic feature encoders, dimensionality-reduction techniques (e.g., PCA, Eigenbrain) compress voxel-level information into tractable latent representations, and ensembles of classical learners (SVM, Random Forest, Naïve Bayes, gradient boosting) discharge the final classification responsibility. Fourth, generalisability — across acquisition protocols, cohorts, and the differential-diagnosis frontier separating AD from vascular and aging-related phenomena — has become the principal axis along which methodological progress is now measured.
Within this consolidated landscape, the proposed Aegis-AD system is positioned to integrate, rather than to replicate, the strongest contributions of the surveyed literature. Specifically, Aegis-AD adopts the deep-to-classic feature-extraction philosophy validated by AlSaeed and Omar (2022) and Salvatore and Nanni (2020); it incorporates the dimensionality-reduction discipline established by Salvatore et al. (2015) and Zhang et al. (2015) to manage voxel-level complexity; it embraces the ensemble decision strategy empirically vindicated by Battineni et al. (2020) and Diogo et al. (2022); and it inherits the multimodal fusion sensibility introduced by Battineni et al. (2021), with explicit provision for the integration of demographic and cognitive covariates alongside imaging-derived features. Equally importantly, the design of Aegis-AD is informed by the clinical caveats articulated by Henry-Feugeas (2007) and Frisoni et al. (2010), which mandate explicit countermeasures against the contamination of the diagnostic signal by physiological aging.
In so doing, Aegis-AD aspires not to introduce a wholly novel paradigm, but rather to articulate a principled architectural synthesis — one that respects the anatomical priors of clinical neuroradiology, the statistical lessons of two decades of supervised and semi-supervised learning, and the representational power of modern deep networks — and to instantiate that synthesis in a reproducible, generalisable, and clinically deployable software system.
End of Literature Review.