ADTNet: Attention-Guided U-Net with Dynamic CNN and Transformers for Skin Cancer Detection

Skin cancer is one of the most common malignancies worldwide, with rising incidence rates emphasizing the need for effective diagnostic and treatment strategies. This study presents the Attention-Guided U-Net, which incorporates Dynamic Convolution and Transformers (ADTNet), a novel deep-learning architecture specifically designed for skin cancer segmentation. ADTNet leverages the strengths of dynamic convolution and transformers within a U-Net framework, enhancing the model’s ability to accurately detect cancerous regions in skin images. A key innovation of ADTNet is its integration of attentionguided mechanisms in the skip connections and transformer modules located at the bottleneck of the architecture. Dynamic convolution improves feature extraction by adjusting to the spatial properties of input images, facilitating more detailed and accurate segmentation of skin lesions. Through comprehensive experimentation, we developed a four-stage version of ADTNet, which achieves an optimal balance between segmentation accuracy and computational efficiency. This version yielded a Dice score of $\mathbf{9 2. 4 \%}$ and an IoU of $\mathbf{8 7. 2 \%}$, demonstrating robust performance in distinguishing cancerous regions from healthy tissue. These metrics indicate the model’s potential for practical implementation in real-world clinical scenarios, where accurate and efficient segmentation is necessary for effective diagnosis.