Dual Concatenated DenseNet with Attention Fusion: A Framework for Skin Lesion Classification Incorporating Multiple Augmentation Techniques and Transfer Learning

Accurate classification of skin lesions is crucial for early detection of skin cancer and effective treatment. However, existing approaches face challenges such as class imbalance, overfitting, and limitations of transfer learning. This study proposes a novel framework, termed Dual Concatenated DenseNet with Attention Fusion (DCDAF), which addresses these challenges through a multi-pronged strategy. The dataset is augmented to mitigate class imbalance while ensuring a completely unseen test set. The DCDAF architecture incorporates multiple customized CNN blocks with varying filter sizes, interspersed with attention modules leveraging channel attention, squeeze-and-excitation attention, and soft attention mechanisms. These attention components enable the model to focus selectively on relevant features. Furthermore, ensemble techniques, including weighted averaging, evidential fusion, and majority voting, are employed to combine predictions from different attention modules, enhancing generalization. Extensive experiments on the HAM10000 dataset demonstrate the efficacy of the proposed approach, achieving an accuracy of 90.24%, outperforming existing methods. The novel DCDAF framework offers a robust solution for skin lesion classification, addressing key challenges and paving the way for improved early detection and diagnosis.