Enhancing skin lesion classification with advanced deep learning ensemble models: a path towards accurate medical...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Skin cancer, particularly malignant melanoma, presents a growing global health concern with increasing incidence rates.
• Early detection of skin lesions is critical for improving patient survival rates and treatment efficacy.

Purpose of the Study:

• To develop and evaluate a deep learning-based system for accurate skin lesion classification.
• To address challenges in skin lesion analysis, including limited data, class imbalance, and image noise.

Main Methods:

• Utilized deep neural networks (ResNeXt101, SeResNeXt101, ResNet152V2, DenseNet201, GoogLeNet, Xception) optimized with Stochastic Gradient Descent (SGD).
• Employed image inpainting for noise reduction and data augmentation to enhance dataset diversity.
• Implemented ensemble techniques (average and weighted average models) with grid search for optimal weight distribution.

Main Results:

• Ensemble models significantly outperformed individual deep learning models in skin lesion classification.
• The average ensemble model achieved a macro-average ROC-AUC of 96%, balancing precision, F1 score, and recall.
• The weighted ensemble model achieved a macro-average ROC-AUC of 97%, demonstrating superior precision and Matthews Correlation Coefficient (MCC).

Conclusions:

• Ensemble techniques effectively enhance the accuracy and robustness of skin lesion classification systems.
• The developed deep learning models show significant potential for improving medical diagnostics in dermatology.
• This approach contributes to better patient outcomes through more reliable early detection of skin cancer.