Hybrid deep feature fusion and ensemble learning for multi-class skin lesion classification

Area of Science:

• Dermatology
• Computer Science
• Artificial Intelligence

Background:

• Automated skin lesion classification faces challenges like high inter-class similarity, intra-class variance, and extreme class imbalance in dermoscopic datasets.
• Existing deep learning models struggle with these inherent dataset complexities.

Purpose of the Study:

• To propose a novel Hybrid Bi-Feature Network (HB-Net) model for enhanced automated skin lesion classification.
• To overcome limitations of existing methods by integrating deep feature extraction with machine learning classifiers.

Main Methods:

• Fine-tuning DenseNet201 and ResNet50 to extract deep features from dermoscopic images.
• Concatenating global average pooled features and fully connected embeddings, followed by Principal Component Analysis (PCA) and Synthetic Minority Over-sampling Technique (SMOTE).
• Classifying refined features using Support Vector Machines (SVM) and a stacking ensemble classifier (Random Forest and Extreme Gradient Boosting).

Main Results:

• The HB-Net model achieved 0.9049 accuracy, 0.8427 recall, 0.9743 specificity, 0.8594 precision, and 0.8489 F1-score on the HAM10000 dataset.
• Demonstrated superior classification performance compared to separate deep learning models on multi-class skin lesion datasets.
• Validation on the ISIC 2019 dataset confirmed the model's effectiveness.

Conclusions:

• The proposed HB-Net model effectively addresses challenges in skin lesion classification.
• Integrating complementary deep features with ensemble machine learning classifiers is a promising approach for reliable skin lesion diagnosis.
• The hybrid framework offers improved performance for automated skin lesion classification.