.png){kind=icon}
.png){kind=icon}
Return
Cross-coupled parallel transformer encoder for biological image segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.bspc.2025.109241.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
With the rising incidence of organ diseases, precise and automated medical image segmentation has become increasingly vital. CNN excels in medical image segmentation has limited capacity for capturing global information. Recent studies use Transformer to encode long-range dependencies, enhancing global feature extraction. However, in the case of medical image segmentation, global as well as local features are required to enhance the precision of segmentation. As a result, this research proposes a method for segmenting medical images that combines Transformers and convolutional networks. Parallel Transformer and CNN encoders were used to capture global and local features, and a residual spatial attention and channel attention fusion module was proposed to better integrate global and local information. During the skip connection process, this study provides a cross feature mixing module that uses convolution and cross attention paths to effectively fuse the features of the encoder and decoder. In addition, we also propose a hybrid atrous convolution of parallel and series cross. This kind of cross-connected atrous convolution fuses features of different branches and different levels to promote the reuse of features and information transfer, so as to realize the effective capture of multi-scale information. We conducted tests using the Synapse multi organic dataset, aortic vessels tree dataset, and Automated Cardiac Diagnosis Challenge dataset. The DICE value on the synapse dataset is 82.71. The DICE value on the AVT dataset is 87.97. The DICE value in ACDC data is 90.83. We have validated the effectiveness of our method by comparing it with existing methods.
Journal
IF:
4.9
Papers:
9.8K
Citations:
2.4W
Organization
No organization information available