Table 3.
References of skin cancer classification with typical CNN frameworks.
| Ref. | Dataset | CNN Architecture | Highlights | Limitations | Performance |
|---|---|---|---|---|---|
| (76) | Self-collected dataset | Deep Belief Network, SVM | By combining deep belief networks and SVM classifiers to handle skin cancer diagnosis tasks with limited datasets, as well as outliers and erroneous data. | The generalization ability of the model is limited. | Accuracy: 0.89 |
| (77) | Self-collected dataset | Resnet-34, ResNet-50 ResNet-101 and ResNet-152 |
Proposed how to improve deep learning-based dermoscopy classification and dataset creation. | Data from more modalities, such as the patient’s medical history, information on other symptoms, are not considered. | Accuracy: 0.85 |
| (78) | Online repositories and the Stanford University Medical Center | Inception-v3 | Used a CNN framework to train a large-scale skin disease dataset and achieve superior results on par with dermatologists. The method was also developed for mobile devices. | More research is required to assess its performance in clinical practice. At the same time, this method is limited to some extent by the amount of data. | Accuracy: 0.6375 (avg.) |
| (79) | MED-NODE | Deep CNN | Compared with previous methods, it directly used CNN to automatically extract features for skin disease images, also had a higher classification accuracy. | Due to the large noise interference of clinical images, there are still some misclassifications. | Accuracy: 0.81 PPV: 0.75, NPV: 0.86 |
| (71) | ISIC-2016 | VGG-16 | It reduces the training time of the model by using the transfer learning strategy while obtaining higher sensitivity and precision. | It is prone to overfitting due to the limited amount of training images. | Accuracy: 0.813 Sensitivity: 0.787 |
| (80) | ISIC-2017, IAD | Inception-v2 | Introducing sonification into the diagnosis of skin cancer lesions to improve the sensitivity of the model. | Differences in the diagnosis of pathologists can affect the prediction results of the model. | AUC: 0.976 Sensitivity: 0.86 Specificity: 0.91 |
| (27) | ISIC-2017 | DenseNet, Dual Path Nets Inception-v4, Inception-ResNet-v2MobileNetV2, PNASNet, ResNet SENet, Xception |
By analyzing 13 factors from 9 different models, they systematically evaluated the factors influencing the choice of CNN structure. | The dataset used in this article is too limited, and it only focuses on the melanoma classification task. | Top accuracy: 0.827 |
| (81) | IAD | VGG-19 | Adopted VGG-19 network to evaluate the thickness of melanoma for the first time. | There are no more pre-training methods utilized for comparison, and precisely predicting melanoma thickness would be more clinically significant. | Accuracy: 0.872 Specificity: 0.840 |
| (82) | Derm7pt | Inception-v3 | A multi-task network was designed to classify the seven-point checklist and skin disease diagnosis. Different loss functions were also designed to handle different input modalities, such as clinical and dermoscopic images, and patient diagnostic results. | Some criteria of the 7-point checklist are unable to be distinguished. | Accuracy: 0.737 |
| (60) | HAM10000 | Deep CNN models | Proposed a method combining CNN with one-versus-all (OVA) for skin disease classification. | The model has not been tested on datasets from various domains and may have a large variance. | Accuracy: 0.929 |
| (83) | HAM10000 ISIC-2019 |
ResNeXt, SeResNeXt, DenseNet Xception, and ResNet |
Adopted a grid search strategy to find the best ensemble learning methods for skin cancer classification. | The amount of training data is still insufficient, and most of models employed in ensemble learning are from the same network architecture. | Accuracy: 0.88 F1 score: 0.89 |