DPE-BoTNeT: Dual Position Encoding Bottleneck Transformer Network for Skin Lesion Classification

Skin cancer is the most common cancer worldwide, and therein malignant melanoma may lead to less than 5-year life expectancy. Early detection and recognition of skin lesion type greatly affect proper treatment to increase the patient’s survival rate. With the progress of various imaging modalities, automatic skin lesion recognition has attracted substantial research attention, and recent deep learning methods using the existing network architectures such as VGGNet, ResNet have demonstrated remarkable performance gain. This study aims to present a novel unified convolution and transformer network, called bottleneck transformer network, for simultaneously modeling local interaction and global de-pendency and further exploits a dual learnable position encoding module to enhance the bottleneck transformer’s position modeling capability. We implement our bottleneck trans-former on the baseline DenseNet and conduct experiments on two benchmark datasets: HAM10000 and ISIC2017. We evaluate that our proposed method outperforms the state-of-the-art deep learning methods.