W-Net: Parallel Encoder Medical Image Segmentation Network Based on CNN and Transformer

Prompt detection of skin cancer is paramount, with image segmentation remaining an essential aspect in clinical diagnosis. At present, CNN models have made remarkable progress in this field, but their convolutional structure leads to limitation of receptive field, which is difficult to effectively embrace extensive dependencies in images, and thus affects the final segmentation accuracy. To address this challenge, researchers have started utilizing models based on the Transformer architecture. This type of model, by incorporating the Transformer and leveraging the self-attention mechanism, can effectively capture long-term image dependencies and global context information by using the self-attention mechanism, however, its capability to grasp localized details remains restricted. Therefore, we propose a parallel encoder structure using two parallel encoder: CNN encoder and Transformer encoder. To combine the feature maps from both encoders, we introduced a module for feature fusion employing a cross-self-attention mechanism, which can effectively capture the interdependence between different encoder features and enhance the semantic information represented by the features. We crafted a wavelet convolution block infused with self-attention to better apprehend fine details and edge characteristics, thereby enhancing the model's proficiency in processing high-frequency data. Our model, as introduced in this study, excels at segmenting the ISIC 2018 dataset. It exhibits augmented precision and robust stability, offering promising support for skin cancer diagnostics and therapy.