Refine Diagnostic Accuracy of Skin Cancer with A Hybrid Deep Network

With the rapid development of computer-aided diagnosis technology, many efforts have been made to refine the early diagnostic accuracy on skin cancer so as to improve the patient’s survival rate. In this work, a hybrid deep network on the basis of a dedicated combination of Resnet-50 and CrossViT is proposed in this research field. Compared to recent baselines, the proposed network shows its strengths in two aspects: 1) the feature representation ability is significantly enhanced in this case by engaging two sub-networks with deconvolution layers, enabling better classification performance with less image detail and information loss; 2) it further boosts the diagnostic accuracy with a distinguishable and stable feature fusion process via adding dense layers in CrossViT, aiming to ensure unimpeded gradient propagation between network layers. Extensive experiments show that the proposed model can outperform recent baselines regarding classification accuracy on diverse types of skin cancer.