In recent years, cancer has been a very undeniable global health challenge. There are various cancer types, such as lymphoma, leukemia, breast cancer [17], lung cancer, skin cancer, and so on. Early skin cancer detection significantly impacts prognosis, and various techniques have been exploited. From histopathological examination to advanced imaging modalities [18], the quest for optimizing predictive models has given rise to the integration of ensemble techniques. This literature review delves into the multifaceted landscape of skin cancer detection methodologies, focusing on the evolving role of ensemble techniques in enhancing survival prediction accuracy. On the other hand, ML is greatly contributing to anomaly detection in various fields, including health care, vehicular networks [19,20], the Internet of Things (IoT), E-commerce, and so on. ML and DL algorithms significantly aid in identifying skin cancer, with early detection potentially leading to successful treatment, making melanoma a significant health concern [21,22,23,24]. Various ML and DL techniques have been applied in existing literature, such as in [25], where the authors presented a convolutional neural network (CNN) based DL stacked ensemble framework for melanoma skin cancer detection using transfer learning. The model uses multiple CNN sub-models and a meta-learner to predict malignant melanoma moles. The model achieves a high accuracy of 95.76%, precision of 95.60%, recall of 96.67%, specificity of 94.67%, F1 score of 94.67%, and area under the curve (AUC) of 0.957% identifying both benign and malignant melanoma. Although this research is important, it could not achieve better accuracy. Similarly, another work in [26] proposed a DL-based skin cancer detection system on an imbalanced dataset. The authors employed the MNIST: HAM10000 dataset that contains seven classes of skin lesions. In order to classify the skin cancer, the authors utilized AlexNet, InceptionV3, and RegNetY-320 techniques. However, the achieved accuracy (91%), F1-score (88.1%), and ROC curve (95%) reflect a poor accuracy as compared to our proposed study.

Moreover, the authors in [27] proposed a CNN-based skin cancer detection system using a publicly available dataset, HAM10000, that includes seven skin cancer types. The authors achieved the following: accuracy (86%), precision (84%), recall (86%), and F-1 score (86%). Thus, all the achieved results fall in the 80s, which reflects the poor performance of the proposed study. Authors in [28] employed a CNN-based approach using a HAM10000 dataset that comprises 6705 benign and 1113 malignant samples and 2197 unknown lesion samples. The proposed model achieved an accuracy of 93.16% on training and 91.93% on testing. Moreover, the authors balanced the dataset of both classes, resulting in an enhanced accuracy of categorization. Despite training several transfer learning models on the same dataset, the outcomes did not surpass those of their proposed model. Another similar work in [29] proposed a CNN-based skin cancer diagnosis that is evaluated using the ISIC 2019 dataset. This work is based on multiclassification system that classifies the cancer types including benign keratosis, melanoma, melanocytic nevi, and basal cell carcinoma. The achieved results depicted an accuracy of 96.91%, which is inefficient as compared to our proposed study. Similar to our work, authors in [9] studied three immune-related mRNAs (SUCO, BTN3A1, and TBC1D2) linked to melanoma prognosis. This study used univariate Cox regression and Kaplan–Meier analysis to compare the overall survival probability between high-risk and low-risk groups, analyzing the time-dependent ROC curve. However, the accuracy of various classifiers is lower as compared to our achieved results.

Furthermore, reference [10] developed a combination of ML and DL-based tools to predict the short-term survival of cutaneous malignant melanoma (CMM), a common malignancy. The study found that additional clinical variables such as sex, tumor site, histotype, growth phase, and age were significantly linked to overall survival, with DNN and RF models showing the best prognostic performance with an accuracy of 91% and 88%, respectively. Reference [30] analyzed mRNA expressions of m5C regulators in colorectal cancer tissues and identified high mutation frequency. NOP2 and YBX1 were highly expressed in prostate, gallbladder, lung, and renal cancers. NSUN6 functions as a tumor suppressor in pancreatic cancer. UV radiation was identified as the primary environmental driver. The authors in [31] trained a HAM10000 ISIC dataset using DL for multiclass skin cancer diagnosis. The proposed model detects the skin lesion with an accuracy of 96.26%.