A Transfer Learning and Feature Ranking-based Feature Extraction Approach for Enhancing Skin Lesion Classification

This study investigates skin lesion classification through feature fusion, focusing on transfer learning-based extraction for improved model discernment. Utilizing VGG16, ResNet, and EfficientNet B0, the research ranks features using methods like model weights, RFE, correlation-based techniques, LASSO regression, gradient boosting, and variance thresholding. The approach aims to enhance diagnostic precision by combining diverse features essential for accurate skin lesion classification. Assessing HAM 10000 and BCN 20000 datasets, the study evaluates the impact of ranking on model performance. Results consistently demonstrate that ranked feature sets outperform initial sets across classifiers (KNN, SVM, CNN) for both datasets. Notably, CNN excels with ranked feature sets from RFE applied to fused features from transfer learning networks.