Table 1.
Show the comparison comparison of lung disease classification models using chest X-ray images.
| Author name | Dataset | Used algorithm | Dataset size | Accuracy | Advantages | Disadvantages |
|---|---|---|---|---|---|---|
| Sanida et al.26 | Chest X-ray images (multi-class) | Custom CNN with VGG19 architecture | 21,165 X-ray images | 98.88% | High diagnostic accuracy with advanced feature extraction and data augmentation | Dataset imbalance required augmentation; computational cost due to architectural complexity |
| Sunnetci and Alkan26 | Chest X-ray images (binary & multi-class) | Bag of Features + Multiple Classifiers | ~ 10,000 images | 99.63% | High accuracy using a biphasic majority voting method, user-friendly GUI | Limited generalization to unseen data; reliance on predefined feature extraction methods |
| Shamrat et al.26 | Chest X-ray images | Fine-tuned MobileNetV2 (MobileLungNetV2) | 112,120 X-ray images (ChestX-ray14) | 96.97% | High classification accuracy; effective use of CLAHE and Gaussian filters for pre-processing; Grad-CAM for visual explanations | Computational complexity; requires high computational power due to fine-tuning and large dataset size |
| Elhanashi et al.26 | Chest X-ray images | Deep learning (multi-classification and ensemble object detection) | Large dataset (not specified in detail) | Promising results (accuracy not provided) | Enables both classification and localization; employs ensemble models for improved detection; applicable to multiple diseases, including COVID-19 | Dataset size and computational requirements not detailed; results compared qualitatively, lacking numerical specificity |
| Holste et al.26 | Chest X-ray images | Long-tailed, multi-label classification models | 350,000 CXRs | Not specified | long-tailed and multi-label challenges; includes 26 clinical findings; large-scale benchmark dataset released for research | Accuracy metrics not detailed; solutions proposed but lack performance quantification for individual models |
| Jeong et al.26 | Chest X-ray images | Optimized ensemble framework for multi-label, long-tailed classification | MIMIC-CXR-LT dataset (size not specified) | Not specified | Addresses multi-label and long-tailed distribution; employs data augmentation and optimized ensemble models; ranks high in the CXR-LT competition | Specific accuracy metrics not provided; focused on competition performance, limiting insights for general applications |
| Sharma et al.26 | Chest X-ray images | Transfer learning-based deep learning | 286 images (rotated through 120° or 140° with 25 augmentations) | Not explicitly stated; combination of models displayed the highest prediction accuracy | Efficient classification of COVID-19 and other infectious diseases; utilizes transfer learning and image augmentation | Small initial dataset size; dependent on image augmentation and transfer learning to generalize |
| Ravi et al.26 | Chest X-ray images | Multichannel EfficientNet-based stacking ensemble | Not explicitly stated | 98% for pediatric pneumonia, 99% for TB, 98% for COVID-19 | Robust and generalizable; employs multichannel models (EfficientNetB0, B1, B2); stacked ensemble improves performance | Lacks dataset size details; computationally intensive due to multichannel and stacked ensemble architecture |
| Kim et al.26 | Chest X-ray images | Transfer learning with EfficientNet v2-M | NIH dataset (unspecified size) and SCH dataset (unspecified size) | 82.15% (NIH), 82.20% (SCH); per-class accuracy: 63.60%-89.90% | End-to-end learning approach; good sensitivity (81.40%) and specificity (up to 94.48%) | Lower testing accuracy for certain classes; lack of dataset size details; variable performance across classes |
| Malik et al.26 | Chest X-rays of COVID-19, pneumothorax, pneumonia, lung cancer, and tuberculosis | Publicly available benchmark datasets | Convolutional Neural Network (CDC Net) with residual network and dilated convolution | 99.39% |
Highly accurate for multi-class classification Incorporates residual networks and dilated convolutions Outperforms pre-trained models (e.g., VGG-19, ResNet-50, Inception v3) |
Relies on publicly available datasets, which might not represent real-world diversity Potential limitations in generalization due to dataset specifics |
| Alshmrani et al.26 | Chest X-rays for COVID-19, Pneumonia, Lung Cancer, Tuberculosis, Lung Opacity, and Normal | VGG19 + Convolutional Neural Network (CNN) | 3615 COVID-19, 6012 Lung Opacity, 5870 Pneumonia, 20,000 Lung Cancer, 1400 Tuberculosis, 10,192 Normal (total ~ 47,089 images) | 96.48% |
Combines pre-trained VGG19 with additional CNN layers High precision and AUC (99.82%) Superior performance in multi-class classification |
Imbalanced dataset sizes may impact generalization Performance depends heavily on the quality of pre-trained model |