Table 1.
State-of-the-art studies to diagnosis PSLs by using pretrained TL models.
| Ref. | Description | * Classification | Dataset | Augment? | Results% | Limitations |
|---|---|---|---|---|---|---|
| [8] | Pigmented Skin Lesions (PSLs) classify into seven classes by using TL approach. | AlexNet | ISIC2018 | Yes | ACC.: 89.7 | Evaluated on single dataset, classify seven-classes, but computational expensive, classifier overfitting. |
| [10] | The combine work of TL and DL. | MobileNet V2 and LSTM | HAM10000 | Yes | ACC: 85 | Classes imbalance problem, binary classification, tested on signal dataset and computational expensive. |
| [11] | The approach employs FixCaps for dermoscopic image classification. | FixCaps | HAM10000 | Yes | Acc.: 96.49 | Classes imbalance problem, evaluated on single dataset, classify only two-classes, and computational expensive. |
| [12] | The approach is a multiclass EfficientNet TL classifier to recognize different classes of PSLs. |
EfficientNet | HAM10000 | Yes | ACC.: 87.91 | Single dataset, accuracy is not good, which limits the detection accuracy, 6 classes only, overfitting and computational expensive |
| [13] | They developed another technique based on CNN and non-linear activations functions. | Employ Linear and nonlinear activation functions either the hidden layers or output layers | PH2 ISIC |
Yes | ACC: 97.5 | Evaluated on two datasets, classify only two-classes, classifier overfitting, and Computational expensive. |
| [14] | CNN model along with activation functions. | Multiple-CNN models | HAM10000 Dataset |
Yes | ACC: 97.85 | Three classes of PSLs and reduced hyper-parameters so computational expensive, classifier overfitting and used only single dataset. |
| [15] | The approach is based on convolutional neural network model based on deep learning (DCNN to accurately classify the malignant skin lesions | DCNN | HAM10000 Dataset |
Yes | ACC:91.3 | Classes imbalance problem, evaluated on single dataset, classify seven-classes, and computational expensive, classifier overfitting. |
| [16] | Differentiate only benign and Malignant. | VGG16 | Skin Cancer: Malignant vs. Benign 1 | Yes | ACC: 89 | Classes imbalance problem, binary classification only two-classes, not generalize solution, and computational expensive. |
| [17] | The approach of the classifier is based on Deep Learning Algorithm | ResNet-152 | ASAN Edinburgh Hallym |
Yes | ACC: 96 | Classes imbalance problem, evaluated on single dataset, classify only two-classes, and computational expensive. |
| [18] | This technique uses features fusion approach to recognize PSLs. | PDFFEM | ISBI 2016, ISIC 2017, and PH2. | Yes | ISBI 2016 ACC:99.8 |
Image processing, handcrafted-based feature extraction approach, which limits the detection accuracy, 6 classes only, classifier underfitting, and computational expensive |
| [19] | Different classifiers are utilized to evaluate the approach. | DT, KNN, LR and LAD | HAM10000 | Yes | ACC:95.18 | Classes imbalance problem, evaluated on single dataset, classify seven-classes, and computational expensive, classifier overfitting. |
| [20] | The approach is a Heterogeneous of Deep CNN Features Fusion and Reduction | SVM, KNN and NN | PH2, ISBI 2016, ISBI 2017 |
Yes | ACC: 95.1% | Image processing, handcrafted-based feature extraction approach, which limits the detection accuracy, 6 classes only and computational expensive |
| [22] | Segmentation and Classification of Melanoma and Nevis | KNN, CNN | ISIC, DermNet NZ | Yes | -- | Classes imbalance problem, evaluated on two datasets, classify only two-classes, and computational expensive. |
| [24] | Segmentation and classification approach | Pretrain CN, moth flame optimization (IMFO) | ISBI 2016, ISBI 2017, ISIC 2018, and PH2, HAM10000 |
Yes | ACC: 91% | Image processing, handcrafted-based feature extraction approach, which limits the detection accuracy, 6 classes only and computational expensive |
| [25] | Dermo-Deep is developed for classification based on two classes | five-layer pretrained CNN architecture | ISBI 2016, ISBI 2017, ISIC 2018, and PH2, HAM10000 |
No | ACC: 96% | Classes imbalance problem, classify seven-classes, and computational expensive, classifier overfitting. |
| [26] | Classification of seven classes to recognize PSLs | Google’s Inception-v3 | HAM10000 | Yes | ACC: 90% | Classes imbalance problem, binary classification only two-classes, and computational expensive. |
| [27] | A DCNN model is developed, which was constructed with several layers, various filter sizes, and fewer filters and parameters | DCNN | ISIC-17, ISIC-18, ISIC-19 | Yes | ACC: 94% | Classes imbalance problem, classify only two-classes, and computational expensive. |
| [29] | Different pretrained models based on transfer learning techniques were evaluated in recent study | DenseNet201 | ISIC | Yes | --- | Limits the detection accuracy, 6 classes only and computational expensive |
* DT: decision tree, KNN: k-nearest neighbor, PDFFEM: pigmented deep fused features extraction method, LDA: Linear Discriminant Analysis, SVM: support vector machine, NN: neural network, ResNet: residual network, DCNN: deep convolutional neural network, FixCaps: improved capsule network, LSTM: long short-term memory, 1 https://www.kaggle.com/datasets/fanconic/skin-cancer-malignant-vs-benign (accessed on 2 September 2022).