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

Automated classification of skin lesion remains a challenging task due to high inter-class similarity, intra-class variance and extreme class imbalance in the dermoscopic datasets. To overcome these limitations, the proposed study presents a Hybrid Bi-Feature Network model as HB-Net, in which the convolutional neural networks are fine-tuned to extract features and subsequently machine learning classifier are used for skin lesion classification.

DenseNet201 and ResNet50 are both fully trainable deep architectures independently trained on the dermoscopic images to find discriminative representations. Global average pooled features are concatenated with fully connected embeddings from both networks to form a unified deep feature vector. Principal component analysis is employed to minimise redundancy and improve the separability of classes, and Synthetic Minority Over-sampling Technique is then used to address the problem of class imbalance. The refined features are further categorized with the help of support vector machine as well as with the stacking ensemble classifier consisting of the random forest and extreme gradient boosting models.

The proposed HB-Net model was evaluated on the HAM10000 dataset, achieving a classification accuracy of 0.9049, recall of 0.8427, specificity of 0.9743, precision of 0.8594, and an F1-score of 0.8489. Experimental findings on a multi-class dataset of skin lesions prove that the proposed hybrid framework achieved better classification performance score than separate deep learning models.

The proposed model is also validated on ISIC 2019 dataset. The findings validate that the integration of complementary deep features with ensemble machine learning classifiers are effective in addressing reliable skin lesion classification.