UCFA‐Net: A U‐shaped cross‐fusion network with attention mechanism for enhanced polyp segmentation

Abstract Enhancing the precision of computer‐assisted polyp segmentation and delineation during colonoscopies assists in the removal of potentially precancerous tissue, thus reducing the risk of malignant transformation. Most of the current medical segmentation models use the traditional U‐shaped network structure, but they suffer from the problem of information loss during the encoding and decoding of images. To advance towards an autonomous model for detailed polyp segmentation, the authors propose a new framework for polyp segmentation called U‐shaped cross‐fusion network with attention mechanism (UCFA‐Net), which employs a pyramid vision transformer as encoder to extract image features at multiple scales. Furthermore, the multi‐scale cross‐fusion module cross‐fuses the different scale features and then goes through the multi‐scale convolutional parallel feedforward transformer module for modelling the global and local information. Finally, progressive attentional up‐sampling module acts as a decoder for up‐sampling with progressive attention to get the final polyp segmentation result. The authors comprehensive testing demonstrates that their network achieves superior average scores across the five datasets and exhibits greater robustness in the face of diverse and demanding scenarios, when compared to current state‐of‐the‐art approaches.