TranSpeckle: An edge‐protected transformer for medical ultrasound image despeckling
Y. Chen, Zhitao Guo
Abstract
Abstract The transformer, a type of neural architecture, has demonstrated exceptional performance improvements in vision and natural language tasks. While overcoming the disadvantages of limited perceptual field and non‐adaptive input content exhibited in CNNs, the computational complexity of the Transformer model increases quadratically with spatial resolution. As such, this model is not frequently employed in image processing tasks such as image denoising, and there is a shortage of studies that investigate ultrasonic image multiplication speckle removal. In light of this, we present TranSpeckle, an effective and efficient despeckle architecture that employs Multi‐Dconv Head Transposed Attention and Dconv Feed‐Forward Network as the core components of its Transformer block. Multiple Transformer blocks are then utilized to implement a hierarchical encoder‐decoder network. TranSpeckle architecture considerably reduces the computational complexity of feature maps while also effectively capturing long‐range pixel interactions and local context information. In this study, an edge protection module is combined to augment the edges of ultrasound images. The module incorporates extracted image edge features into the TranSpeckle architecture, which ameliorates the issue of edge information loss engendered by the image despeckling process. Extensive experimental results clearly show that our proposed network outperforms state‐of‐the‐art methods in terms of quantitative metrics and visual quality.