Multi-Class Skin Cancer Detection Using CNN-Architecture Based Deep Learning Models

Skin diseases are a global health issue that can cause physical discomfort and emotional distress. Manual diagnosis is time-consuming and subjective, relying heavily on dermatologists’ expertise. This study proposes a comprehensive approach to address this by leveraging deep learning models for skin disease diagnosis. The HAM10000 dataset, comprising over 10,015 dermoscopic images labelled with seven different diagnostic categories, serves as a benchmark for evaluating skin disease prediction models. The proposed work applied three models, namely, CNN, GoogleNet, and ResNet to the HAM10000 dataset. The class imbalance issue of this dataset is handled by mean and median sampling strategy. Performance will be evaluated using standard metrics like accuracy, precision, recall, and F1-score. In addition, a thorough analysis is conducted on the AUC-ROC curves to prove the efficacy of the best performing model. The study provides a detailed methodology, experimental setup, results, and conclusions, summarizing fndings and potential implications for clinical practice. It is found that the CNN model with mean sampling strategy has the best outcome in terms of stable prediction.