MRI-based Alzheimer’s Classification Gains Precision Through Hybrid Feature Selection
A new computational framework for Alzheimer’s disease classification combines transfer learning, hand-tuned image preprocessing, and machine learning postprocessing to separate Alzheimer’s disease, mild cognitive impairment, and cognitively normal MRI scans with high retrospective accuracy. Built on the ADNI1 Complete 1Yr 1.5T dataset, the pipeline processed 2,294 MRI observations from 639 subjects and was designed to reduce one of the most common sources of inflated performance in imaging studies: data leakage between scans from the same person.
The model paired VGG19 feature extraction with Harris Hawks Optimization and simulated annealing, then passed the selected features into a random forest classifier. In the reported test setting, the system reached 98.70% accuracy, along with 98.55% macro-precision, 98.93% macro-recall, 99.28% macro-specificity, 98.73% macro-F1, and 99.40% macro-ROC-AUC. The study also compared its best accuracy against a previously reported benchmark of 98.20%, a difference of 0.50 percentage points.
Why the Pipeline Matters for Three-Class Dementia Imaging
Three-way classification is harder than the binary Alzheimer’s-versus-control problem that dominates much of the literature. Mild cognitive impairment sits in the gray zone between normal aging and overt dementia, and MRI features can be subtle, noisy, and highly redundant. A classifier that performs well on one dataset can fail quickly when the input distribution shifts, especially if preprocessing is inconsistent or if the same subject’s images leak into both training and test sets.
This framework attacked the problem at several layers. Spatial and intensity normalization were used first, followed by histogram equalization and double bilateral filtering to clean the images before deep feature extraction. Those steps are not cosmetic. In structural brain MRI, contrast differences, scanner variability, and local noise can obscure the atrophy patterns a model needs to recognize. By standardizing the images before feeding them into VGG19, the pipeline aimed to make downstream features less sensitive to acquisition artifacts and more responsive to biologically relevant structure.
The choice of VGG19 also reflects a practical reality in neuroimaging AI: transfer learning can extract useful representations even when dataset size is modest compared with the demands of training a deep network from scratch. Rather than relying on end-to-end learning alone, the model used a hybrid architecture in which deep features were selected and then compressed through optimization before classification. The design is computationally conservative, but it also raises the ceiling on performance when the number of subjects is limited.
Subject-Level Splitting and the Risk of Inflated Performance
The strongest part of the methods is the subject-level partitioning. Each patient’s scans were kept together to minimize leakage, which matters because MRI slices or repeated observations from the same person can make a model appear better than it really is. In dementia imaging, this is not a minor technical detail. A classifier that memorizes subject-specific anatomy can look spectacular on paper while offering little genuine generalization.
Here, the reported metrics suggest the model handled the three-class problem with high internal consistency. Macro-level scores are especially relevant because they treat Alzheimer’s disease, MCI, and cognitively normal groups more evenly than accuracy alone. In a clinical context, missing MCI or mislabeling early Alzheimer’s disease has different consequences than misclassifying a clear control, so balanced performance measures provide a more honest picture of system behavior.
One sentence captures the key result: the classifier separated all three cognitive states with near-ceiling performance on retrospective ADNI data.
Optimization and Random Forest as the Final Decision Layer
The most distinctive element of the workflow is the hybrid Harris Hawks Optimization with simulated annealing. Feature selection is often the hidden determinant of success in radiology AI, because deep networks can produce a large number of correlated descriptors that add noise without adding diagnostic value. HHO-SA attempts to narrow the feature set to the most informative subset, and simulated annealing adds a search strategy that can escape poor local minima.
The downstream random forest then serves as a robust decision layer. In image classification tasks, random forests can be attractive because they tolerate non-linear relationships, handle mixed feature importance well, and reduce the overfitting that may emerge when too many descriptors are carried forward from a deep network. The combination suggests a deliberate tradeoff: let the convolutional backbone learn rich features, then use optimization and ensemble learning to strip away excess complexity before classification.
The result is a pipeline that is technically compact but methodologically layered. For applied neuroimaging research, that matters because many proposed deep learning systems are difficult to reproduce, rely on specialized training configurations, or collapse when moved to new data. A framework built in Python and PyTorch, with transparent preprocessing and conventional machine learning at the end, is easier to inspect and potentially easier to adapt.
What the Metrics Support, and What They Do Not
The reported performance numbers are strong enough to attract attention, but they remain confined to a retrospective benchmark dataset. ADNI is a valuable resource, yet it is not the same as real-world clinical deployment, where scanner protocols vary, comorbidities are common, and MRI quality is less controlled. A model that performs at 98.70% accuracy on curated data still needs external validation before it can be considered clinically reliable.
The study also does not establish whether the classifier would change patient management, reduce diagnostic uncertainty, or improve referral timing in memory clinics. It identifies a computational path toward high-fidelity stratification of Alzheimer’s disease, MCI, and normal cognition, but it does not answer the harder clinical question of how such a tool would fit into workflow alongside neuropsychological testing, biomarker assessment, and specialist review.
Even so, the choice to use de-identified ADNI MRI data without recruiting new participants keeps the analysis squarely in the retrospective research domain, where computational methods can be tested efficiently before moving toward prospective evaluation. No new human data were collected, and no additional institutional ethical approval was required for the analysis as conducted.
Reproducibility Will Decide Clinical Value
The central claim is not that MRI alone can replace clinical diagnosis, but that a carefully engineered image-analysis pipeline can classify cognitive states with unusually high consistency on a standard research dataset. If future studies reproduce the same performance across external cohorts, scanner platforms, and broader patient populations, the approach could become a useful component of dementia imaging research.
For now, the main limitation is the same one that shadows much of the field: strong retrospective accuracy does not guarantee transportability. The next test is whether the hybrid VGG19-HHO-SA-random forest structure holds when the data are less tidy, the case mix is broader, and the disease labels are messier than those in ADNI.