Alzheimer's disease is diagnosed in part from structural changes visible on brain MRI — hippocampal atrophy, ventricular enlargement, cortical thinning — that occur at different spatial scales and different depths of a network's receptive field. No single convolutional block family is built to capture all of that at once, so this work proposes a hybrid CNN combining four block types: residual connections (ResNet) for gradient stability, multi-branch multi-scale convolutions (Inception) for scale-varying anatomical features, depthwise-separable convolutions (MobileNet) for lightweight refinement, and dense connectivity (DenseNet) for feature reuse deep in the network.
Critically, block placement is not a fixed template. A single-block ablation was run first, and the hybrid's layout follows directly from what it showed: Inception is given two stages since it benefits most from depth, ResNet is placed mid-stack as a stabilizer rather than the lead feature extractor, MobileNet is placed late as cheap non-bottleneck refinement, and DenseNet is given the deepest, highest-capacity stage. A component-level leave-one-block-out ablation, an attention variant (CBAM), calibration analysis, and Grad-CAM visualizations round out the evaluation.
Trained from scratch (no pretrained weights) on 44,000 MRI slices across four severity classes, the model reaches 99.38% test accuracy (99.25% ± 0.08% averaged over 5 independent runs) with only 0.785M parameters.
Each block family contributes a different inductive bias to brain MRI feature extraction, combined into one architecture:
Skip connections prevent vanishing gradients as the network deepens — placed mid-stack as a stabilizer, not the lead feature extractor.
Parallel 1×1 / 3×3 / 5×5 branches capture fine cortical texture and coarse ventricular structure at once — given two stages here.
Depthwise-separable convolutions cut parameters ~8–9× versus standard convolutions — used late as shape-preserving refinement.
Every layer receives all preceding feature maps, maximizing feature reuse — given the deepest, highest-capacity stage.
A preprocessing pipeline tailored to consistent, stable training of a batch-normalized CNN from scratch:
Block placement is set by single-block ablation evidence, not a fixed template: Inception is given two stages, ResNet stabilizes mid-stack, MobileNet refines late, DenseNet gets the deepest stage. 0.785M parameters vs. the plain-CNN control's 0.980M — roughly a quarter smaller.
Block placement follows a single-block ablation run beforehand, not a fixed textbook ordering — every stage above has a measured reason.
Roughly a quarter smaller than the plain-CNN control (0.980M) tested in this study, at 0.34 GFLOPs — practical for constrained clinical hardware.
The leave-one-block-out ablation deletes each block outright — no substitute layer is inserted — so removed variants have strictly fewer parameters, the stricter test.
Each block is deleted outright — no substitute layer — so every variant below has fewer parameters than the full model, a stricter test than substitution.
| Variant | Accuracy | Δ vs. Full | Parameters |
|---|---|---|---|
| Hybrid CNN (full, all 4 blocks)Full model | 99.38% | — | 0.785M |
| w/o ResNet | 94.89% | −4.49pp | 0.730M |
| w/o 2nd Inception stage | 98.38% | −1.00pp | 0.649M |
| w/o DenseNet | 99.12% | −0.26pp | 0.263M |
| w/o MobileNet | 99.36% | −0.02pp | 0.732M |
ResNet is the most critical block in context (−4.49pp when removed) despite being the weakest standalone block in isolation — its role as gradient stabilizer matters more inside the combined model than its raw representational power alone.
| Model | Accuracy | Parameters | FLOPs |
|---|---|---|---|
| Hybrid CNN (no attention)Best | 99.38% | 0.785M | 0.340G |
| Hybrid CNN + CBAM | 99.08% | 0.791M | 0.340G |
McNemar's test on paired test-set predictions: χ² = 4.30, p = 0.038 — the −0.30pp gap is statistically significant. Attention did not help this architecture on this task.
| Model | Parameters | FLOPs |
|---|---|---|
| BaseCNN (plain-conv control) | 0.980M | 0.536G |
| CNN + ResNet only | 0.198M | 0.062G |
| CNN + DenseNet only | 0.187M | 0.623G |
| CNN + Inception only | 0.369M | 0.250G |
| CNN + MobileNet only | 0.117M | 0.029G |
| Hybrid CNN (proposed)Proposed | 0.785M | 0.340G |
Single-block variants shown for architectural size comparison only — standalone accuracy for these was not collected in this run.
An interactive Gradio-powered web application that deploys the Hybrid CNN for real-time inference. Choose a CNN architecture, upload a brain MRI slice, and the app returns the predicted probability of each Alzheimer's disease category — MildDemented, ModerateDemented, NonDemented, VeryMildDemented.
Loading Demo...
This work presented a hybrid CNN for four-class Alzheimer's severity classification from brain MRI, with block placement grounded in single-block ablation evidence rather than an arbitrary ordering. Trained from scratch, the model demonstrates that a compact, evidence-driven architecture can reach very high accuracy on this task without pretrained weights or a larger over-parameterized backbone.
The model reached 99.25% ± 0.08% test accuracy across five independent runs, with a well-calibrated confidence profile (ECE 0.0024) and only 0.785M parameters — roughly a quarter smaller than the plain-CNN control tested in this study. A component-level leave-one-block-out ablation showed ResNet — the weakest standalone block — to be the most critical block in context, underscoring that a block's isolated strength does not predict its contribution inside a combined architecture.
A CBAM attention variant, statistical significance testing (McNemar's test, confidence intervals), and Grad-CAM explainability visualizations were used to check the result from multiple angles — not just accuracy, but whether the model's confidence and attention are trustworthy.
Evidence-driven hybrid design — block ordering set by a preceding single-block ablation rather than a fixed template, with the rationale reported alongside the architecture.
True leave-one-block-out ablation — each block is deleted outright, not substituted, so every ablation variant has strictly fewer parameters than the full model, the stricter comparison.
Statistical rigor — repeated-run mean ± std, Wilson confidence intervals, and McNemar's test check whether reported differences are meaningful, not just point estimates.
Calibration and explainability — Expected Calibration Error and Grad-CAM visualizations go beyond accuracy to check whether confidence and attention are trustworthy.
Deployable Gradio application — an end-to-end inference app where a brain MRI slice is uploaded and the model returns per-class probabilities in real time.
Single dataset — trained and evaluated on one Kaggle source; augmented datasets of this kind carry known risk of near-duplicate leakage across splits.
No external validation — not yet tested on an independently sourced cohort (e.g. ADNI, OASIS) to confirm generalization beyond this dataset's acquisition protocol.
Fixed split, single seed for ablations — only the Hybrid CNN result is repeated across 5 seeds; ablation and CBAM comparisons reflect a single training run each.
Attention didn't help here — CBAM's channel + spatial attention gave a small, statistically significant accuracy decrease on this task and architecture.
2D slices, not 3D volumes — the model classifies individual MRI slices rather than reasoning over the full 3D brain volume.
Cross-validation — K-fold or multiple random splits to confirm results aren't an artifact of one particular split.
External cohort validation — held-out evaluation on ADNI or OASIS to test generalization across scanners and protocols.
Robustness testing — accuracy under noise, contrast shift, and minor acquisition variation beyond training augmentation.
Repeated-run ablations — extend the 5-seed protocol to the leave-one-block-out and CBAM comparisons, not just the full model.
3D volumetric extension — moving from 2D slice classification to full 3D volume reasoning for richer spatial context.