Table 1 Literature review summary.
Reference | Techniques Used | Advantages | Disadvantages | Results |
|---|---|---|---|---|
Neelum Noreen et al. (2020)7 | Deep learning, Inception-v3, Xception, Random Forest, SVM, K-Nearest Neighbors | High accuracy, faster diagnosis, decision-support tool for radiologists | Limited explainability, reliance on manual feature engineering, Limited datasets, no external validation | Accuracy: 94.34% |
J. Chathumina et al. (2024)8 | Deep learning, XAI techniques, Comparative analysis | Addresses explainability in DL, Comparative analysis of XAI techniques | Black-box nature of DL models, explainability still evolving, Focuses only on explainability, lacks practical application | Not mentioned |
N. Shamshad et al. (2024)9 | Transfer learning with VGG-16, ResNet-50, and MobileNet on MRI images using CNNs | Enables high accuracy (especially with VGG-16), efficient feature extraction, reduces training time, and improves generalization | ResNet-50 is computationally heavy; MobileNet has lower accuracy and unstable validation; risk of overfitting | VGG-16: 97.2% accuracy, AUC 0.974, 22% processing time; ResNet-50: 96% accuracy, 43% time; MobileNet: 87% accuracy, 35% time |
R. Singh et al. (2024)10 | Ensemble deep learning using ResNet50 and EfficientNet-B7 with dynamic weighted averaging on MRI data | High accuracy, improved generalization, reduced misclassification, dynamic model adaptability | Increased computational complexity, dependent on a single dataset, limited real-time feasibility | Ensemble model accuracy: 99.53%; EfficientNet-B7: 98.20%; ResNet50: 97.4%; ensemble had highest precision and F1-scores |
Ghazanfar Latif et al. (2018)11 | Wavelet feature extraction, Random Forest, Region-growing segmentation | Effective tumor classification, high multiclass accuracy | Limited dataset (35 cases), computationally intensive, Small dataset, lacks external validation | Accuracy: 96.08% |
A.S.M. Shafi et al. (2021)12 | Ensemble learning, ROI extraction, Collewet normalization, SVM, Majority voting | Strong performance in tumor and lesion detection, high accuracy | Complex pre-processing steps, potential for overfitting, Limited testing on other conditions, not generalizable | Accuracy: 97.5% |
Nahid Ferdous Aurna et al. (2022)13 | Two-stage CNN, PCA, Pre-trained models, SVM | Excellent accuracy, real-time validation with UI | Complex model requiring significant computational resources, Limited to specific brain tumor types, lacks multi-institutional validation | Accuracy: 98.51% |
Jose Antonio Marmolejo-Saucedo et al. (2022)14 | CNN, XAI techniques, numgrad-CAM | Improves model explainability, good accuracy | Underperformance after model calibration, high abstention rate, Underperformance in some cases, requires further calibration | Accuracy: 97.11% |
Dieine Estela Bernieri Schiavon et al. (2023)15 | CNN, Hyperparameter optimization, Data augmentation, XAI | High accuracy, improved reliability via XAI techniques | No direct integration for clinical workflows yet, Limited to specific datasets, lacks large-scale validation | Accuracy: 96% |
Eric W. Prince et al. (2023)16 | DL classifier, Bayesian DL, Predictive uncertainty | Improved accuracy with uncertainty calibration, non-invasive diagnosis | Original model overfitted, high abstention rate (34.2%), Limited generalization, requires more diverse data | Accuracy: 95.5% |
Burak Taşcı (2023)17 | Densenet201, gradcam, INCA feature selection, SVM | High accuracy, outperforms SOTA methods | High computational cost, explainability limited to gradcam, Limited to two datasets, needs further validation in clinical settings | Accuracy: 98.65% |
