A robust stacked ensemble strategy with multi-optimizer CNN models for skin cancer classification.

Area of Science:

• Dermatology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Skin cancer is a global health concern requiring accurate diagnostic tools.
• Early detection significantly improves patient survival rates.
• Computer-aided diagnostic systems offer potential for enhanced skin cancer diagnosis.

Purpose of the Study:

• To optimize a stacked ensemble strategy for improved skin lesion classification.
• To evaluate the performance of various convolutional neural network (CNN) architectures and ensemble methods for melanoma detection.

Main Methods:

• Utilized four CNN architectures trained with multiple optimizers (Adam, SGD, RMSprop, AdaMax).
• Implemented and compared four ensemble strategies: soft voting, hard voting, weighted voting, and stacked ensemble with meta-learners (GB, KNN, SVM, LR, NB).
• Evaluated models using accuracy, precision, recall, F1-score, and ROC-AUC on the Melanoma Skin Cancer Dataset.

Main Results:

• The stacked ensemble model with Gradient Boosting (GB) meta-classifier achieved 97.4% validation accuracy.
• Independent test accuracy reached 91.80% with a ROC-AUC of 96.87%.
• Ensemble learning, especially stacked ensembling, demonstrated significant improvements in classification accuracy and reliability.

Conclusions:

• Optimized stacked ensemble strategies significantly enhance the accuracy and reliability of melanoma detection.
• This approach shows promise for improving computer-aided diagnostic systems for skin cancer.
• Further research in ensemble learning can advance early skin cancer diagnosis and patient outcomes.