Table 1. References of skin cancer segmentation with typical CNN frameworks in the literature.
Only the metrics for the best-performing architectures or datasets are presented in studies that involve the comparison of multiple architectures or datasets.
| Ref/Year | Dataset | Architecture | Highlights | Limitations | Performance |
|---|---|---|---|---|---|
| (Ameri, 2020)/2020 | HAM10000 | AlexNet | The suggested approach eliminates the need for complex procedures of lesion segmentation and feature extraction by taking an unprocessed image as input and learning directly from the raw data. | Only 3,400 images from the dataset were used due to the need for an equal number of benign and malignant images for training. | ACC = 0.84, SE = 0.81, SP = 0.88 |
| (Yao et al., 2022)/2022 | ISIC-2017, ISIC-2018, ISIC-2019, 7-PT | RegNetY-3.2GF | The study proposes a novel Multi-weighted New Loss method to address the issue of class imbalance and improve accuracy in detecting key classes such as melanoma. RegNetY performed the best on the ISIC2018 dataset. | Almost all publicly accessible skin disease image datasets suffer a problem of severe data imbalance that might affect the performance of CNNs. | BACC = 0.858 |
| (Perez, Avila & Valle, 2019)/2019 | ISIC-2017 | Inception-ResNet-v2, MobileNetV2 , PNASNet, ResNet , SENet, Xception, VGG16, VGG19, and DenseNet | The authors systematically assessed the factors that impact the selection of a CNN structure by examining 13 different factors across nine models. | The article’s dataset has limitations as it is smaller in size compared to other studies and its exclusive focus on classifying melanoma. | Top-1 ACC = 0.827 |
| (Javid et al., 2023)/2023 | A recompilation of ISIC datasets | ResNet 50, EfficientNet B6, Inception V3, and Xception | The results obtained from each individual model are inputted into a meta-learner in order to combine and utilize the outputs from these models to make a final prediction. | – | ACC = 0.935, SE = 0.9, PR = 0.96, F1 score = 0.92 |
| (Alwakid et al., 2022)/2022 | HAM10000 | Modified version of Resnet-50 | The proposed method suggests utilizing DL to precisely extract a lesion zone. The approach involves enhancing the image quality using ESRGAN and then using segmentation to isolate Regions of Interest (ROI). | To showcase the effectiveness of the proposed technique, it is necessary to conduct more experiments on a sizable and intricate dataset that encompasses potential cancer cases. | ACC = 0.86, SE = 0.86, PR = 0.84, F1 score = 0.86 |
| (Raza et al., 2022)/2022 | Clinical repositories in Korea | InceptionV3, Xception, InceptionResnetV2, DenseNet121, VGG16 | A novel stacked ensemble framework has been introduced in the study, specifically designed to augment generalizability and bolster robustness in the context of acral lentiginous melanoma classification. | – | ACC = 0.979, SE = 0.978, PR = 0.98, F1 score = 0.98 |
| (Gupta & Mesram, 2022)/2022 | ISIC 2016–17 | AlexNet, DenseNet-121 | The study suggests a mixed CNN model that involves merging a pre-trained AlexNet CNN model with an optimized pre-trained DenseNet-121 CNN model. | Numerous healthcare institutions possess substantial patient data; however, they face challenges in making this information accessible to the public due to privacy concerns. | ACC = 0.9065, SE = 0.91, PR = 0.9065, F1 score = 0.91 |
| (Esteva et al., 2017)/2017 | Clinical repositories | GoogleNet, Inception v3 | The model is adapted to be used on mobile devices. It is predicted that by the year 2021, there will be approximately 6.3 billion smartphone subscriptions globally. | Further investigations are necessary to evaluate how this method performs in a clinical setting. | ACC = 0.721 |