A Deep Learning Approach vs Optimized Machine Learning Models: A Comparative in Skin Cancer Classification

Currently, World Health Organization reports evidence that skin cancer has become more relevant in recent decades. Globally, approximately 32,000 cases of melanoma skin cancer and around three million cases of non-melanocytic lesions are reported annually. Timely identification of skin cancer contributes to timely and effective diagnosis and treatment in potential patients. A comparative analysis is presented between a deep learning approach using a custom-designed convolutional neural network and XGBoost and Random Forest models optimized with hyperparameters for skin cancer classification. The HMNIST skin lesion image descriptor dataset was used, and minority classes were balanced through oversampling with induced normal noise in the training dataset, increasing the number of image descriptor data from 18,891,264 to 88,362,288. The comparison was made based on weighted performance analysis of parameters such as accuracy, precision, recall, f1-score, and specificity, as well as hardware performance, weighing training time and percentages of required CPU and RAM. The proposed model showed improvements of up to 6.03% in accuracy, 10.9% in precision, and 8.7% in f1-score in melanoma lesion detection. Additionally, the CPU utilization rate during training improved by 43%. The proposed deep learning model is presented as an alternative for timely identification and recognition of malignant skin cancer lesions.