Table 3 Comparison between this work and previous studies.
From: Hybrid of DSR-GAN and CNN for Alzheimer disease detection based on MRI images
Study | Techniques | Data used | Key findings | Evaluation metrics | Advantages | Weaknesses | Future directions |
|---|---|---|---|---|---|---|---|
Transfer learning with MRI data, AlexNet | Small MRI datasets | Classified AD into four stages | Accuracy: 91.7% across four AD stages | Efficient for small datasets | Low accuracy compared to state-of-the-art | Fine-tuning CNN layers, exploring other architectures, testing larger datasets | |
Ensemble model (VGG-16, EfficientNet-B2) | Imbalanced MRI datasets | Outperformed single models in accuracy and AUC | Accuracy: 97.35%, AUC: 99.64% | Addressed class imbalance, robust classification | Limited to imbalanced datasets | Test with larger datasets | |
CNN-based model (DEMNET) | Kaggle, ADNI datasets | High performance in classifying AD stages | Accuracy: 95.23%, AUC: 97%, Kappa: 0.93 | Class imbalance handled via SMOTE | Lower performance on larger datasets | Robustness testing on ADNI | |
CNN on sMRI images | ADNI dataset | High accuracy, reduced parameters | High accuracy, automated feature extraction | Parameter efficiency | Limited robustness due to dataset constraints | Incorporate patient history | |
ADD-Net, a deep CNN model | MRI scans | Outperformed DenseNet169, VGG19 | Accuracy: 98.63%, AUC: 99.76% | Lightweight model | Limited scalability to other imaging modalities | Incorporate transfer learning | |
Hybrid models (CNN, Bidirectional LSTM) | DementiaBank transcripts | Significant hyperparameter optimization | Accuracy: 93.31%, significant hyperparameter optimization | Achieving significant hyperparameter optimization via GridSearch | Achieved low accuracy | Improve performance on larger datasets | |
LSTM networks | Patient data over time | Predicts AD progression | Outperformed existing approaches in predictions | Focusing on the temporal relationships in patient data | Reached medium testing accuracy | Highlighted temporal relationships | |
Ensemble CNNs, patch-based approach | ADNI, NRCD datasets | Competitive accuracy in AD diagnosis | Accuracy: 85.55% (ADNI), 90.05% (NRCD) | Providing an effective framework | Achieved low testing accuracy | Address overfitting, data scarcity | |
Deep CNN architecture | MRI datasets | Achieved high classification accuracy | Accuracy: 99.05% in classification | High classification performance | High complexity architecture 7,866,819 parameters | Refine preprocessing, multi-modal approaches | |
ResNet50, automatic feature extraction | ADNI dataset | Outperformed traditional methods | Accuracy: 85.7%—99% | Achieved high accuracy | Consume long training time | Automation of diagnosis | |
Custom CNN with data augmentation | MRI dataset | Addressed overfitting, high-accuracy | Accuracy: 99.1%, 99% precision, sensitivity, specificity | High classification performance | Dataset diversity and size were limited | Enrich the dataset with diverse sources | |
Neural network classifier with VGG16 | Two MRI datasets | High performance in early AD diagnosis | Accuracy: 90.4% & 71.1%, AUC: 0.969 & 0.85 | The model was trained on two different datasets | Achieved low accuracy | Expand dataset, improve accuracy | |
Review of deep learning techniques | Various neuroimaging modalities | Identified challenges and methodologies | Identified VGG-16 as a preferred model | effectiveness in large-scale image recognition tasks | Reached low accuracy compared to other pre-trained models | Create comprehensive datasets, enhance model transparency | |
SVM with rule-extraction methods | National Alzheimer’s Coordinating Center dataset | Achieved high F1 scores for classification | F1: 98.9% (binary), 90.7% (multiclass) | Explainability, robust multi-class performance | High computational costs for feature space reduction | Focus on model explainability | |
This work | A hybrid of DSR-GAN and CNN | 6,400 MRI images based on four classes | Achieved high accuracy with a super-resolution technique as an image processing | Accuracy: 99.22%, Precision: 99.01%, Recall: 99.01%, F1-score: 99.01%, AUC: 100%, PSNR: 29.30 dB, SSIM: 0.847, MS-SSIM: 96.39% | Novel hybrid model, super-resolution enhancement | Only 1,700 images dataset for super-resolution training | Large dataset and additional deep learning methodologies, such as Vision Transformer (ViT) |
