EigenGradCAM with AugSmooth: Enhancing Explainability and Precision in Multi-Stage Alzheimer's Disease Detection

Background

• Early identification of mild cognitive impairment matters because the passage describes it as a crucial intervention period, while traditional diagnostic pathways can miss structural differences and confuse MCI with moderate dementia.Traditional pathways include neuropsychological evaluation, cerebrospinal-fluid analysis, and manual neuroimage inspection.Source: §1 · p. 2

Clinical motivation

• Deep-learning models can make accurate diagnoses without showing clinicians which brain regions or image features support them, complicating validation and patient-safety review.The passage notes that incorrect reasoning may go undetected without interpretation tools.Source: Existing Problems · pp. 2–3

Research challenges

• Grad-CAM heatmaps may miss hippocampal and entorhinal regions, gradient maps can change under small input perturbations, and severe class imbalance makes accuracy alone inappropriate.The cited dataset example has 12 Moderate dementia images versus 640 No impairment images, a 50:1 ratio.Source: Existing Problems · p. 3

Research gap

• The stated research gap is the absence of a method combining PCA of gradient-weighted convolutional features with Gaussian smoothing in weight space for four-stage 2D axial MRI staging.The passage also reports that prior AD heatmap validation was qualitative and that the combined approach had not been studied.Source: Research Gap · p. 3

Research aims

• The study aims to classify four AD stages with a custom CNN, use WGAN-GP to address class imbalance, and improve saliency localization and stability through eigenvector-based mapping and gradient-weight perturbation.The approach is intended to be evaluated for both classification performance and heatmap neuroanatomical localization.Source: Research Objectives · p. 3

Contributions

• 97.2% test accuracy, 97.0% balanced accuracy, and 0.944 Matthews correlation coefficient are reported for the proposed approach, with AUC values from 0.96 to 0.99 across four AD stages.The reported heatmaps consistently identify the hippocampus and entorhinal cortex.Source: Research Objectives · p. 4

Epidemiological Context and Clinical Motivation

• The review presents AD as progressing from normal cognition through MCI to dementia, with hippocampal and entorhinal atrophy observable on structural MRI and early MCI described as important for disease-modifying therapy.The passage uses this staging context to motivate multi-stage MRI diagnosis.Source: A. Epidemiological Context and… · p. 5

Deep Learning for Alzheimer's MRI Classification

• Prior CNN studies report high benchmark accuracy, but many use variants of the same Kaggle data, often without addressing Moderate Dementia undersampling or independent evaluation.The cited patient-level clinical sMRI evaluation reported 63% balanced accuracy without class-imbalance mitigation, unlike benchmark reports near 99%.Source: B. Deep Learning for Alzheimer's MRI… · pp. 5–6

Vision Transformers and Attention Architectures

• Transformer and attention studies report strong classification results, while the review identifies anatomical validation as uncommon and notes one example with hippocampal-boundary Dice evaluation.RAE-ViT is reported at 94.2% with an attention-map Dice similarity coefficient of 0.89 against hippocampal boundaries.Source: C. Vision Transformers and Attention… · p. 6

Class Imbalance, Synthetic Augmentation, and GAN-Based Methods

• The reviewed four-class benchmark has a 50:1 No Impairment-to-Moderate Dementia training-image ratio, and prior work has applied WGAN-GP to generate Moderate Dementia MRI samples.The review describes the broader gap as combining WGAN-GP, multi-stage classification, and gradient-attribution explainability.Source: D. Class Imbalance, Synthetic… · pp. 6–7

Explainable AI Foundations and Gradient-Based Saliency Methods

• Grad-CAM variants pursue improved localization, Eigen-CAM applies PCA to raw activations, and SmoothGrad averages maps over noisy input copies; the review says these approaches lack gradient-weight-space noise averaging for neuroimaging.The proposed AugSmooth perturbs Grad-CAM weight vectors before PCA rather than perturbing input pixels.Source: E. Explainable AI Foundations and… · pp. 7–8

XAI Applied to Alzheimer's Disease

• AD imaging explainability studies have identified hippocampal regions, but the review describes many saliency evaluations as qualitative and identifies quantitative faithfulness assessment as an unresolved need.It cites 3D Grad-CAM work identifying the precuneus and hippocampus, while other listed approaches lack insertion/deletion metrics or quantitative faithfulness scores.Source: F. XAI Applied to Alzheimer's Disease · p. 8

Multimodal Fusion, Federated Learning, and Clinical Translation

• Recent work explores MRI-PET fusion and federated learning, while the review frames unimodal 2D axial MRI classification and interpretability assessment as a repeatable milestone for further development.Examples include federated approaches and multimodal systems with reported attention to hippocampal and entorhinal regions.Source: G. Multimodal Fusion, Federated… · p. 9

CNN Architecture Principles Supporting the Proposed Framework

• The reviewed custom CNN design follows literature-backed choices including batch normalization in convolutional layers, dropout in fully connected layers, and Adam optimization.The passage also identifies softmax with categorical cross-entropy and ReLU as the selected multi-class output and activation choices.Source: H. CNN Architecture Principles… · p. 9

Accuracy–Interpretability Trade-off

• 86% of reviewed XAI medical-imaging studies used only qualitative evaluation, while a meta-analysis reported a 5–7% AUC penalty for interpretable models.The passage also notes that clinical usability depends critically on quantitative faithfulness evaluation, which remains largely absent.Source: §I · p. 10

Convergent Failure

• The review identifies a three-part gap: quantitative heatmap faithfulness on neuroanatomically validated targets, class-balanced WGAN-GP training, and PCA-based gradient attribution with weight-space smoothing have rarely been combined.It frames this combination as the contribution of EigenGradCAM with AugSmooth.Source: §I · p. 10

• No published study combined EigenGradCAM’s PCA decomposition of gradient-weighted feature maps with AugSmooth for four-stage AD staging on 2D axial MRI.Source: J. Identified Research Gap · p. 10
• Prior EigenGradCAM applications were limited to general computer-vision benchmarks, while AD heatmap evaluations remained qualitative without insertion/deletion AUC.Source: J. Identified Research Gap · p. 10
• Existing work had not validated whether WGAN-GP-balanced training and PCA-smoothed attribution jointly improve classification accuracy and heatmap clinical alignment.Source: J. Identified Research Gap · p. 10
• The review compares representative frameworks by classifier, XAI method, accuracy, balanced accuracy, quantitative XAI evaluation, and limitations.Source: J. Identified Research Gap · p. 10, Table 1 · p. 10

3. Proposed Methodology

• The framework targets four-class AD staging from 2D axial MRI while addressing Moderate Dementia class imbalance, opaque CNN predictions, and gradient-method instability.Source: §3 · p. 12

3.1 Dataset

• The dataset contains approximately 15,000 MRI images across four clinical stages, with substantially fewer Moderate Demented samples than Non-Demented samples.Source: §3.1 · p. 12

3.2 WGAN-GP Synthetic Balancing

• WGAN-GP synthesizes minority-class MRI samples until each of the four training classes contains 2,560 images.FID is used to assess the distributional similarity of generated and real images.Source: §3.2 · p. 14

3.3 Preprocessing and Data Augmentation

• Images are min-max rescaled to 0–1, and training augmentation applies rotation, zoom, and horizontal flips to vary scan geometry.Source: §3.3 · pp. 14–15

3.4 CNN Architecture

• The custom four-class CNN uses batch normalization after convolutional layers, dropout at rate 0.5 after dense layers, and a softmax output.Source: §3.4 · pp. 15–16

3.5 Training Strategy

• Training runs for at most 100 epochs with batch size 32, stops after 10 validation-loss epochs without improvement, and saves the lowest-loss checkpoint.Source: §3.5 · p. 17

3.6 EigenGradCAM with Augmented Smoothing (AugSmooth)

• EigenGradCAM applies class-specific Grad-CAM weights before PCA, while AugSmooth averages Gaussian-perturbed weights to stabilize principal eigenvectors without changing input images.The perturbation uses σ=0.15 and N=50 samples.Source: §3.6 · p. 18

3.7 Model Evaluation

• Evaluation pairs accuracy with MCC and balanced accuracy for imbalanced four-class assessment, and uses precision-recall analysis where Moderate Impairment has only 12 test samples.Source: §3.7 · pp. 20–21

4.1 Implementation Environment

• The experiments ran in Kaggle Notebooks, with the hardware and software environment documented in Table 4.The complete process, from WGAN-GP balancing through CNN training and heatmap creation, was conducted in one notebook.Source: §4.1 · p. 22, Table 4 · p. 22

4.2 Experimental Configuration

• The model was optimized on 10,240 balanced training images and evaluated only on 1,279 held-out test images.The training set contained 2,560 images per class; the test set was kept separate during training and hyperparameter tuning.Source: §4.2 · p. 23

4.3 Baseline Models for Comparison

• Five literature baselines were evaluated on the same four-class benchmark, but they lack balanced accuracy, Moderate-class sensitivity, and XAI metrics.The authors caution that raw accuracy alone can mislead when the test set has 12 Moderate Impairment samples versus 640 No Impairment samples.Source: §4.3 · p. 23, Table 6 · p. 23

4.4 Evaluation Metrics

• Balanced accuracy and MCC are primary metrics because the test distribution is severely imbalanced, while per-class F1 and one-vs-rest AUC assess stage-specific performance.Precision-recall analysis is also reported because it is more informative than ROC analysis alone for the 12-sample Moderate Impairment class.Source: §4.4 · p. 24

5.1 Training and Validation Performance

• 98.4% training accuracy and 97.2% validation accuracy were reported, with a 1.2-point difference and final test loss of 0.123.The authors report smoothly converging accuracy curves and decreasing training and validation losses.Source: §5.1 · p. 25, Figure 5 · p. 25

5.2 Test Set Classification Performance

• 97.2% test accuracy, 97.0% BAS, and 0.944 MCC were achieved on the held-out test set of 1,279 images.The test set included only 12 Moderate Impairment cases, compared with 640 No Impairment cases.Source: §5.2 · p. 25, Table 7 · p. 25

5.3 Confusion Matrix Analysis

• The confusion-matrix analysis reports mostly correct classifications, with errors concentrated between adjacent stages and most Moderate Impairment cases classified correctly.The authors relate the most frequent error, Very Mild classified as No Impairment, to structural similarity in 2D axial MRI.Source: §5.3 · pp. 26–27, Figure 6 · p. 26

5.4 ROC Curve and AUC Analysis

• The macro-average AUC is 0.975, with AUCs of 0.99 for No Impairment and 0.96 for Moderate Impairment.The Moderate Impairment estimate is based on 12 test samples.Source: Table 8 · p. 27, Figure 7 · p. 27, §5.4 · p. 28

5.5 EigenGradCAM with AugSmooth Heatmap Results

• Heatmaps associate AD and Mild Impairment activations with the hippocampus and entorhinal cortex, while No Impairment activations are more widespread.AugSmooth uses Gaussian noise with σ = 0.15 over N = 50 perturbations of Grad-CAM weight vectors before PCA, producing visually smoother maps.Source: §5.5 · pp. 28–29, Figure 8 · p. 28, Table 9 · p. 29

5.6 Comparison with Baseline Models

• 97.2% test accuracy is lower than the raw accuracies reported by five baselines, while the framework also reports 97.0% BAS and 0.944 MCC and uniquely combines balanced training with XAI evaluation.The authors caution that baseline raw accuracies are difficult to interpret without balanced accuracy or MCC on the imbalanced test set.Source: Table 10 · p. 30, §5.6 · p. 30

6.1 Overview of the XAI Pipeline

• EigenGradCAM with AugSmooth combines class-specific Grad-CAM weights, PCA of gradient-weighted feature maps, and Gaussian perturbation of weights before decomposition.The approach averages 50 perturbed realizations with σ = 0.15 and overlays the heatmaps on MRI images with blend factor α = 0.4.Source: §6.1 · p. 31

6.2 Heatmap Visualization per AD Stage

• Activations are widespread for No Impairment, more diffuse for Very Mild Impairment, and localized in hippocampal and entorhinal regions for Mild and Moderate Impairment.The authors relate these stage-dependent patterns to the expected progression of AD-related atrophy and to regions identified in prior work [6][52].Source: §6.2 · p. 31

6.3 Effect of AugSmooth on Heatmap Stability

• Compared with standard Grad-CAM and EigenGradCAM without AugSmooth, AugSmooth is reported to reduce residual instability and produce smoother, spatially coherent heatmaps.The Gaussian perturbations are applied to Grad-CAM weight vectors rather than input pixels, leaving the input MRI distribution unchanged.Source: §6.3 · pp. 31–32, Table 11 · p. 32

6.4 Neuroanatomical Alignment and Clinical Relevance

• Neuroanatomical alignment is assessed qualitatively by visual inspection against known landmarks, not quantitative faithfulness metrics or expert boundary annotations.The authors identify insertion/deletion AUC and neuroradiologist annotations as priorities for future work.Source: §6.4 · pp. 33–34, Table 12 · p. 33

7.1 Interpretation of Classification Performance

• 97.2% test accuracy, 97.0% BAS, and 0.944 MCC were reported; baseline accuracy-only results of 99.6%–99.92% are not directly comparable.The authors note that the baselines provide neither BAS nor MCC and use a test distribution in which No Impairment is 50.0% and Moderate Impairment 0.9%.Source: §7.1 · p. 35

7.2 Significance of WGAN-GP Class Balancing

• WGAN-GP balancing produced an F1-score of 0.91 for Moderate Impairment after equalizing training classes at 2,560 images each.The discussion describes a 50:1 No Impairment-to-Moderate Impairment training imbalance and says unbalanced training would suppress Moderate Impairment sensitivity.Source: §7.2 · p. 35

7.3 Interpretation of Explainability Results

• Heatmaps show activation shifting from diffuse cortical regions in No Impairment to focused hippocampal and entorhinal regions in Mild and Moderate Impairment.The authors report that AugSmooth suppresses gradient-shatter artifacts without altering the input-image distribution.Source: §7.3 · pp. 35–36

7.4 Limitations

• The evaluation is limited to 2D axial 128x128 slices and one dataset, with no quantitative faithfulness metrics or neuroradiologist validation.The Moderate Impairment test set has 12 instances, and the dataset does not specify patient gender ratios, precluding study of gender-based classification bias.Source: §7.4 · p. 36

7.5 Future Work

• Proposed next steps include 3D MRI and independent ADNI/OASIS validation, quantitative faithfulness measures, multimodal MRI/PET fusion, and federated learning.The authors also propose neuroradiologist hippocampal annotations to enable Dice similarity coefficient evaluation.Source: §7.5 · p. 36

8. Conclusion

• 97.2% test accuracy, 97.0% Balanced Accuracy Score, and 0.944 Matthews Correlation Coefficient were achieved, with AUC values from 0.96 to 0.99 across four stages.The conclusion attributes these benchmark results to training on a balanced distribution of 2,560 images per class.Source: §8 · p. 37
• EigenGradCAM applies class-discriminative Grad-CAM weighting before PCA, while AugSmooth perturbs activation weight vectors to produce stable, spatially coherent heatmaps.The heatmaps identify the hippocampus and entorhinal cortex as most activated for Mild and Moderate Impairment, consistent with the cited Braak staging pathogenesis.Source: §8 · p. 37
• The approach remains limited to 2D axial slices, one dataset, and qualitative XAI faithfulness evaluation.The conclusion identifies 3D MRI, independent ADNI/OASIS validation, and quantitative faithfulness evaluation as directions toward practical clinical applications.Source: §8 · p. 37