Table 1.
State-of-the-art comparison.
| Ref. | Approach | Dataset | Features inputs | Methodology | Comparison parameter | Performance metrics | Limitations |
|---|---|---|---|---|---|---|---|
| Hosny and Kassem 4 | Refined residual deep convolutional network |
ISIC 2017 dataset |
Image features and 2018 |
Refined residual deep convolutional network |
Accuracy, Sensitivity, Specificity, F1-score, AUC-ROC |
Accuracy: 0.94–0.98, Sensitivity: 0.90–0.97, Specificity: 0.95–0.98, F1-score: 0.94–0.98, AUC-ROC: 0.97–0.99 |
Limited dataset, lack of diversity in skin types, limited comparison with other methods |
| Bukhari et al. 8 | Multi-parallel depthwise separable and dilated convolutions with Swish activations |
ISIC 2018 dataset |
Image features | Multi-parallel depth-wise separable and dilated convolutions with Swish activations |
Dice similarity coefficient, Jaccard similarity coefficient, Sensitivity, Specificity, Accuracy |
Dice similarity coefficient: 0.882, Jaccard similarity coefficient: 0.788, Sensitivity: 0.892, Specificity: 0.980, Accuracy: 0.964 |
Limited dataset, lack of diversity in skin types |
| Mustafa et al. 19 | ANN with color and texture features |
DermQuest and PH2 Dataset |
Color and texture features |
ANN with color and texture features |
Accuracy, Sensitivity, Specificity |
Accuracy: 93.9%, Sensitivity: 91.2%, Specificity: 96.4% |
Limited dataset, limited evaluation metrics |
| Lingaraj et al. 21 | support vector machine |
Dermoscopic Images |
Color, Texture, Shape, and Statistical Features |
Feature extraction using Gabor filter and SVM model |
Accuracy, Sensitivity | Accuracy: 83.25%, Sensitivity: 88.41%, Specificity: 78.04% |
Small dataset size, performance not compared with other state-of-the-art techniques |
| Khan et al. 22 | Convolutional neural network |
Dermoscopic Images |
Transfer Learning (Inception-V3, VGG-19, ResNet-50) |
Preprocessing and CNN-based model |
Accuracy, Sensitivity | Accuracy: 88.8%–91.2%, Sensitivity: 89.2%–92.8% |
Limited dataset size, lack of comparison with other with other techniques |
| Liang and Wu 23 | Convolutional neural network |
Dermoscopic Images |
Patch-based features |
Preprocessing and CNN-based model |
Dice Coefficient, Sensitivity |
Dice Coefficient: 91.91%, Sensitivity: 91.93% |
Small dataset size, lack of comparison with other state-of-the-art techniques |
| Ashraf et al. 24 | Artificial neural networks |
Dermoscopic mages |
Color and texture features |
Feature extraction using Gabor filter and ANN model |
Accuracy, Sensitivity | Accuracy: 96.5%, Sensitivity: 96.9% |
Limited dataset size, performance not compared with other state-of- the-art techniques |
| Ashraf et al. 25 | Deep Stacked Patched Auto-Encoders |
Dermoscopic Images |
Patch-based features |
Feature extraction using DSAE and SVM model |
Accuracy, Sensitivity | Accuracy: 92.5%, Sensitivity: 93.5% |
Small dataset size, lack of comparison with other state-of-the-art techniques, limited explanation of feature extraction process |
| Ali et al. 26 | Deep Residual Network |
Dermoscopic Images |
Texture and Shape Features |
Feature extraction using DRN and SVM model |
Accuracy, Sensitivity, | Accuracy: 94.12%, Sensitivity: 94.98% |
Small dataset size, lack of comparison with other state-of-the-art techniques |
| Jeyakumar et al. 27 | CNN, VGG16, ResNet50, DenseNet169, InceptionResNetV2 |
ISIC 2018 dataset |
Image features |
CNN, VGG16, ResNet50, DenseNet169, InceptionResNetV2 |
Accuracy, Sensitivity, Specificity, F1-score, AUC-ROC |
Accuracy: 0.917–0.952, Sensitivity: 0.889–0.917, Specificity: 0.960–0.988, F1-score: 0.894–0.929, AUC-ROC: 0.957–0.985 |
Lack of diversity in skin types, limited evaluation metrics |
CNN: convolutional neural network; SVM: support vector machine; ANN: artificial neural network.