Table 8 Comparison with other related works.
References | Dataset | Method | Precision | Recall | AUC | Accuracy |
|---|---|---|---|---|---|---|
HAM10000 | Features extracted from the pre-trained MobileNetV3 were optimised using Artificial Rabbits optimisation algorithm based on Gaussian mutation and crossover operator | 0.9002 | 0.8965 | – | 88.71 | |
HAM10000 | UNet was used to segment the skin lesion image and the features were extracted from the segmented images using hybrid optimisation algorithm by combining Bacterial Foraging Optimization (BFO) and Particle Swarm Optimization for classification with the deep CNN | 96.02 | 95.37 | 0.98 | 96.46 | |
HAM10000 | A Deep Belief Network has been proposed for the classification of skin lesion using improved Electromagnetic Field Optimization algorithm | 94.99 | – | – | 85.99 | |
HAM10000 | Convolutional capsule layers were added in the Capsule Neural network for the multi-class classification of skin lesions | 95.37 | – | – | 95.24 | |
HAM10000 | Wavelet transform and ResNet101 model has been used to extract the discriminatory features using wavelet transformation, pooling, and normalization for skin lesion classification | 95.84 | – | – | 95.73 | |
HAM10000 | High dimension contrast transforms for segmentation of skin lesions and then the features extracted from the pre-trained DenseNet201 using segmented images were classified using Extreme Learning Machine | – | – | – | 88.39 | |
HAM10000 | They have combined discrete wavelet transform with CNN architecture for the classification of skin cancerous lesion images | 94.00 | 91.00 | – | 94.00 | |
HAM10000 | They have proposed an ensemble of five different CNNs combined using weighted average ensemble method | – | – | 96.00 | – | |
ISIC2017 | Features were extracted from the segmented lesion images using Region Average Pooling method and classification using a Linear Classifier | 60.7 | – | 0.842 | 83.00 | |
ISIC2017 | Full resolution convolutional network for segmentation of lesion from the dermoscopic images and the classification of segmented images using deep CNNs | – | 75.33 | 81.57 | ||
ISIC2017 | Handcrafted and deep features were extracted from the segmented lesion image and classified using support vector machine classifier | – | – | – | 85.3 | |
ISIC2017 | Pre-trained NASNetMobile model has been fine-tuned for skin lesion classification task | 81.77 | – | – | 82.00 | |
ISIC2017 | Color, shape and texture features were extracted from the skin lesion image and classified using k-Nearest Neighbour classifier | – | 80.00 | 0.68 | – | |
ISIC2017 | Two dimensional Fourier transformed image is passed to a customized CNN for feature extraction | – | – | 0.93 | 83.00 | |
Kaggle dataset | Deep features have been concatenated and K-best features were selected and classified using XGBoost classifier | 85.5 | – | – | 85.91 | |
Proposed method | HAM10000 ISIC-2017 | Dual autoencoder networks have been used one for the extraction of the spatial features and other for the extraction of frequency domain features; the multi-level attention has been integrated in the encoder part; the combined features were classified using Extreme Learning Machine | 97.68 86.75 | 97.59 86.62 | 0.98 0.95 | 97.66 86.68 |
