Skin Cancer Classification Using a Hybrid Pre-trained CNN with Random Forest Classifier

Abstract

Skin cancer, which causes approximately 10 million deaths annually worldwide, is projected to see a rapid increase in cases if early diagnosis is not achieved. Traditional diagnostic methods, relying visual examination and histopathology, are often subjective and time-consuming. Recent advancements in Convolutional Neural Networks (CNN) have shown promise in automating and enhancing the accuracy of image analysis for the early detection of skin cancer. Current CNN approaches have leveraged transfer learning and hybrid models to improve performance. Nonetheless, the potential for overfitting remains, and there is still room for enhancing model accuracy. This study investigates the potential of pre-trained CNN models—such as DenseNet-201, InceptionV3, MobileNet, ResNet50, and VGG16—by modifying these models to improve their ability to differentiate between malignant and benign skin lesions. Additionally, a hybrid model approach is introduced, concatenating extracted features from various modified pre-trained CNN architectures and processing them through machine learning classifiers. The modifications and evaluations revealed that the proposed models surpassed the performance of state-of-the-art CNN models on ISBI 2016 datasets. The enhanced models achieved an impressive accuracy rate of 94.20%, marking a significant improvement over traditional CNN models and underscoring the potential of advanced CNN techniques in improving skin cancer diagnosis outcomes.