Table 5.

Summary of works on deep learning models in skin cancer diagnosis.

Reference	Skin Cancer Category	Deep Learning Model	Description of Approach Used	Dataset	Key Contribution	Limitations	Performance Evaluation Metrics and Results
[72]	Melanoma skin cancer	Recurrent neural networks	Classification phases uses modified deep learning algorithm by coalescing optimization concepts from RNNs	PH2	Superior to existing algorithms in terms of optimal segmentation and classification for melanoma skin cancer	Heavy dependence on parameters for segmentation and classification	Algorithmic analysis including specificity: 0.94915, sensitivity: 0.83051, precision: 0.89091, F1-score: 0.85965, etc.
[76]	Skin cancer detection	Autoencoders	Dataset is reconstructed using autoencoder model, reconstruction and spiking networks contribute to enhanced performance	ISIC	Feature sets obtained from convolution model are suitable for merging	Model extracts many unnecessary and irrelevant features	Specificity: 0.9332, sensitivity: 0.9372, precision: 0.9450, F1-score: 0.9411, accuracy: 0.9354
[81]	Skin cancer diagnosis	Long short-term memory model	Tumor marker data values were used to train and test an LSTM model	Two independent medical centers	LSTM model demonstrates superiority while dealing with irregular data and can be used when time intervals between tests vary	Inability to analyze irregular tumor marker data for cancer screening	Time-to-cancer diagnosis in different risk groups, risk stratification
[87]	Binary classification, multi-class skin cancer diagnosis	Deep neural network	CNN architectures trained on large datasets and evaluated against algorithm-assisted clinicians’ results	Edinburgh and SNU datasets	Model serves as an ancillary tool to enhance diagnostic accuracy of clinicians	Outcome of algorithm is significantly affected by composition of input images; performance is sub-optimal if input image quality is low	Improvement in sensitivity and specificity by 12.1% and 1.1%, respectively
[88]	Malignant tumor detection	Deep belief network	Analyze patient data from deep learning perspective, merged with patient attributes and case reports to construct an expert system helping to predict the probability of early cancer	Jiangsu Provincial Hospital of Traditional Chinese Medicine	Relatively effective dimensional reduction and noise cancellation technique, reduces missed clinician diagnoses during endoscopy and treatment	Medium runtime in comparison to other deep learning methods	Accuracy: 0.8148, precision: 0.8571, recall: 0.6, F1 score: 0.7059
[91]	Melanoma, carcinoma, keratosis	Deep convolutional neural network	Classifies skin cancer using ECOC SVM and deep CNN, images are cropped to reduce noise	Pretrained on ImageNet, Internet Images for fine-tuning	Multi-class skin cancer classification using fine-tuned pretrained ImageNet model	Model does not extend to ABCD (asymmetry, border, color, diameter) rule	Accuracy: 0.942, specificity: 0.9074, sensitivity: 0.9783
[96]	Tumor causing somatic mutations	Deep Boltzmann machine	Multi-modal deep Boltzmann machine approach for prediction of somatic mutation genes that undergo malignant transformation, model learns relation between germline and mutation profiles using data	-	Genome-based diagnostic test to monitor for the presence of cancer-driving mutations	Sample size of is limited, Whole Exome Sequencing (WES) data displayed at gene level	Average accuracy: 0.7176, p-value
[99]	Melanoma skin cancer	Extreme learning machine	After pre-processing, Otsu method is employed to segment region of interest, subsequently, feature extraction is applied to mine important characteristics, deep belief network is used to categorize and classify	ISIC for training, SIIM-ISIC melanoma for validation	Optimized Pipeline feature designed for efficient detection of melanoma from images, DBN uses Thermal Exchange Optimization Algorithm as new meta-heuristic method	Computationally very intensive and time consuming	Accuracy: 0.9265, specificity: 0.8970, sensitivity: 0.9118, PPV: 0.8676, NPV: 0.9412