A Framework for Skin Cancer Classification using Ensemble Transfer Learning and Data Fusion

Abstract

Skin cancer is a serious and growing health problem worldwide. Its cases keep rising every year. Early diagnosis plays a key role in improving survival. Manual diagnosis through visual checks and dermoscopy depends on the doctor's experience. It can also give inconsistent results. Biopsy is the most reliable method. But it is invasive and time-consuming. These limits have pushed researchers toward deep learning-based systems. Most past studies use only one model. This lowers accuracy when skin lesions vary in colour, texture, and shape. To fix this, this study proposes an ensemble model combining EfficientNetB3 and Xception. The ISIC 2016 and ISIC 2017 datasets were combined for more diverse training data. Class imbalance was fixed using SMOTE. Both models were trained using transfer learning with ImageNet weights. They were trained under 3-fold cross-validation on a fixed test set. Mixup augmentation was also used during training. Predictions from both models were combined using equal-weighted averaging. The model was evaluated using accuracy, precision, recall, ROC-AUC, and confusion matrices. The ensemble reached a test accuracy of 95.22%, outperforming several existing methods. The ensemble also reached a sensitivity of 92.57%, a specificity of 97.88%, and a ROC-AUC of 0.9878. It improved over the standalone Xception model, which reached 95.01% accuracy, by 0.21 percentage points. It also improved over the standalone EfficientNetB3 model, which reached 94.90% accuracy, by 0.32 percentage points. These results confirm that this ensemble approach, with SMOTE-based balancing, offers a reliable solution for automated skin cancer classification.