Fostering trust and interpretability in skin cancer classification: a hybrid framework of deep learning and machine learning with explainable AI

This study introduces a novel hybrid approach (namely, Conv2D-RF, which stands for “2D convolutional layers with a random forest head”) for improved skin cancer detection and classification, combining image processing with a Convolutional Neural Network (CNN) architecture alongside a Random Forest (RF) classifier. The primary preprocessing steps include addressing class imbalance through oversampling and data augmentation, as well as normalizing the input images. The CNN extracts features from images, which the RF then utilizes for classification purposes. Two widely recognized datasets, namely ISIC and HAM10000, were utilized. Our proposed Conv2D-RF model performed the best on both datasets, with accuracy scores of 97% on ISIC and 99% on HAM10000. Our objective includes improving the skin cancer classification model’s reliability and interpretability, so we use the Gradient-weighted Class Activation Mapping (Grad-CAM) technique to understand our hybrid model’s decision-making process.