Skin Cancer Classification in Dermatological Images based on a Dense Hybrid Algorithm

A tool for identification and classification of skin cancer type in dermatoscopic images based on a dense hybrid algorithm is presented. Noise reduction steps in body hair are proposed using image processing techniques, and features are extracted from the HAM10000 dataset using a DenseNet121 network, where the Principal Component Analysis technique is applied to reduce 1,367 to 374 components, and subjected to classification using XGBoost. The developed hybrid dense algorithm is compared with computational learning algorithms by weighting accuracy, precision, F1 score, ROC and AUC as software comparison metrics, and training time, RAM usage and CPU requirement as hardware validation methods. Improvements of up to 3.9% in accuracy, 4% and 4.9% in precision, and 4.9% and 3.3% in F1 score are found for the carcinoma and melanoma classes. In addition, the percentage of memory used in processing is improved by up to 4.09 times. The developed dense hybrid algorithm presents robustness in performance tests, and is presented as a viable alternative in the identification of cancer-like diseases in skin lesions.