Abstract Methodology Architecture Results Demo Authors Conclusion ⭐ GitHub
Deep Learning for Medical Imaging · 2026

NeuroFusion-HCNN: A Multi-Architectural Hybrid Convolutional Neural Network for Multi-Class Alzheimer’s Disease Classification from MRI Images

Tanjin Adnan Abir

99.38% Test Accuracy
99.99% ROC-AUC (macro)
44,000 MRI Slices, 4 Classes
0.785M Parameters

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.

Medical Imaging Convolutional Neural Network Alzheimer's Disease Brain MRI Ablation Study Attention (CBAM) Explainability PyTorch

End-to-End Pipeline

1
Data
44,000 T1-weighted MRI slices, 4 classes
2
Split
70 / 15 / 15 train / val / test, configurable
3
Single-Block Ablation
Each block tested alone to inform layout
4
Hybrid Design
4 blocks combined per ablation evidence
5
Model Training
From scratch, AdamW, cosine schedule
6
Evaluation
99.38% accuracy · ablations · Grad-CAM

Hybrid Block Design

Each block family contributes a different inductive bias to brain MRI feature extraction, combined into one architecture:

🔗 ResNet

Skip connections prevent vanishing gradients as the network deepens — placed mid-stack as a stabilizer, not the lead feature extractor.

x → conv → conv → + x

🔀 Inception

Parallel 1×1 / 3×3 / 5×5 branches capture fine cortical texture and coarse ventricular structure at once — given two stages here.

x → [1×1, 3×3, 5×5, pool] → concat

📱 MobileNet

Depthwise-separable convolutions cut parameters ~8–9× versus standard convolutions — used late as shape-preserving refinement.

3×3 depthwise → 1×1 pointwise

🧬 DenseNet

Every layer receives all preceding feature maps, maximizing feature reuse — given the deepest, highest-capacity stage.

x₀, x₁, x₂, x₃ → concat → x₄

Image Preprocessing

A preprocessing pipeline tailored to consistent, stable training of a batch-normalized CNN from scratch:

1
Resize — every MRI slice resized to 128 × 128 pixels
2
Random horizontal flip — p = 0.5, training split only
3
Random rotation — ±10°, training split only
4
Color jitter — brightness/contrast ±0.1, training split only
5
Tensor conversion — to a 3-channel float tensor
6
Normalize — ImageNet mean/std, for stable batch normalization from scratch
MRI slice (128×128, uint8)
↓ after preprocessing
tensor [3, 128, 128], normalized

Dataset Statistics

Class Distribution
MildDemented (smallest) 10,000
NonDemented (largest) 12,800
📊 Max/min class ratio 1.28× — close to balanced across 4 classes
Train / Val / Test Split
70/15/15 Train / Val / Test
Train 70% · 30,800
Validation 15% · 6,600
Test 15% · 6,600
Configurable in one place · 60/20/20 and 50/25/25 also supported

Hybrid CNN Architecture

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.

Output
Softmax Output 4 classes
Classifier Head Global pool, dropout, FC
Feature Compression
Transition Layer 4×4, 208ch
Dense Block, 6 Layers 8×8, 416ch
Refinement
MobileNet Block 8×8, 224ch
Multi-Scale Feature Extraction
Inception Block 2 + Pool 8×8, 224ch
ResNet Block, Stride 2 16×16, 160ch
Inception Block 1 + Pool 32×32, 96ch
Stem
Conv 3×3, Stride 2 + BN + ReLU 64×64, 48ch
Input MRI Slice 128×128, 3ch

Key Design Decisions

🎯

Evidence, Not a Template

Block placement follows a single-block ablation run beforehand, not a fixed textbook ordering — every stage above has a measured reason.

0.785M Parameters

Roughly a quarter smaller than the plain-CNN control (0.980M) tested in this study, at 0.34 GFLOPs — practical for constrained clinical hardware.

🧠

True Block Removal

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.

Hyperparameters

Input resolution128 × 128
Batch size64
OptimizerAdamW
Learning rate1e-3, cosine
Weight decay1e-4
Dropout0.3
Max epochs50
Early stop patience8
Train / val / test70 / 15 / 15%
Parameters0.785M
FLOPs0.340G
HardwareTesla T4 GPU
Output4, Softmax

Experimental Results

99.38%
Test Accuracy (Best Run)
99.25%
Accuracy, Mean over 5 Runs
99.99%
ROC-AUC (Macro OvR)
Hybrid CNN · Mean over 5 Runs 1 / 3
Accuracy99.25% ± 0.08%
Precision (Weighted)99.26%
Recall (Weighted)99.25%
F1-Score (Weighted)99.25%
ROC-AUC (Macro OvR)99.99%
Calibration Score (1 − ECE)99.76%
Test Set · 6,600 Samples 2 / 3

Rows = True Class · Columns = Predicted Class

Mild Moderate NonDem VeryMild
Mild 1475 0 1 3
Moderate 0 1470 0 0
NonDem 1 0 1912 15
VeryMild 4 0 17 1702
6,559
Correct
41
Errors
0.62%
Error Rate

Most common error: VeryMildDemented → NonDemented (17 cases, 0.99% of that class) — the clinically hardest boundary in the dataset.

Per-Class Recall · Best Run 3 / 3
🟢 NonDemented Support: 1,928
Recall
99.17%
🔵 VeryMildDemented Support: 1,723
Recall
98.78%
🟡 MildDemented Support: 1,479
Recall
99.73%
🔴 ModerateDemented Support: 1,470
Recall
100.00%

Component-Level Ablation — Leave-One-Block-Out

Each block is deleted outright — no substitute layer — so every variant below has fewer parameters than the full model, a stricter test than substitution.

VariantAccuracyΔ vs. FullParameters
Hybrid CNN (full, all 4 blocks)Full model99.38%0.785M
w/o ResNet94.89%−4.49pp0.730M
w/o 2nd Inception stage98.38%−1.00pp0.649M
w/o DenseNet99.12%−0.26pp0.263M
w/o MobileNet99.36%−0.02pp0.732M

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.

Attention Ablation — CBAM Channel + Spatial Attention

ModelAccuracyParametersFLOPs
Hybrid CNN (no attention)Best99.38%0.785M0.340G
Hybrid CNN + CBAM99.08%0.791M0.340G

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 Complexity — Architecture Only

ModelParametersFLOPs
BaseCNN (plain-conv control)0.980M0.536G
CNN + ResNet only0.198M0.062G
CNN + DenseNet only0.187M0.623G
CNN + Inception only0.369M0.250G
CNN + MobileNet only0.117M0.029G
Hybrid CNN (proposed)Proposed0.785M0.340G

Single-block variants shown for architectural size comparison only — standalone accuracy for these was not collected in this run.

Alzheimer's MRI Classifier

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...

Research Team

TAA
Tanjin Adnan Abir
M.Sc. in Applied Physics & Electronics
Jahangirnagar University, Savar, Dhaka

Summary & Future Directions

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.

Key Contributions

1

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.

2

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.

3

Statistical rigor — repeated-run mean ± std, Wilson confidence intervals, and McNemar's test check whether reported differences are meaningful, not just point estimates.

4

Calibration and explainability — Expected Calibration Error and Grad-CAM visualizations go beyond accuracy to check whether confidence and attention are trustworthy.

5

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.

⚠️

Limitations

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.

🔭

Future Directions

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.

"The weakest standalone block was the most critical one in context — a reminder that ablation questions asked in isolation and in combination can have different answers."

— Hybrid CNN for Alzheimer's Disease MRI Classification, 2026

Read Abstract View Results Try the Demo