Enhancing Skin Cancer Classification with Ensemble Models

Diagnosis in the early stages of skin cancer is often delayed due to high similarity among different types of skin lesions. This paper compares deep learning and machine learning models for the classification of skin lesions. However, the training dataset consists of unbalanced and rare skin disease entities, which poses challenges in automatically classifying skin cancer. Additionally, the model’s cross-domain adaptability and robustness are important considerations. Deep learning methods have recently gained popularity in skin cancer classification, offering satisfactory results and addressing the aforementioned issues. The dataset comprises images depicting benign and malignant skin lesions. The deep learning models utilized in this paper are Mobilenet, DenseNet121, Xception, ResNet50, VGG19, and ResNet152. The machine learning models employed are SVM, Logistic Regression, and KNN. The ABCD features capture asymmetry, edge irregularity, colour variation, and diameter, while the LBPH descriptors represent texture information. The evaluation metrics employed in this research are accuracy, precision, recall, and F1 score. The findings indicate that the ensemble of deep learning models achieves a 95% accuracy and performs well in other metrics. SVM and KNN also exhibit promising results, highlighting the effectiveness of ABCD and LBPH features in machine learning. This research offers valuable insights into the classification of skin lesions, benefiting dermatologists and healthcare professionals.