| U-Net | UNet++［15］ | A dense jump connection strategy is used to introduce a deep supervision mechanism to motivate the network to better learn and utilize hierarchical features | Deeper and multi-scale feature extraction and fusion are realized to improve segmentation performance | Large number of model parameters and high computational complexity |
| Attention U-Net［46］ | AttentionGate attention mechanism is used to eliminate noise and irrelevant information of jump connections and extract key features | It effectively captures important local and global information in the image and improves the accuracy and performance of image segmentation | Model segmentation performance is improved to a lesser extent |
| Transformer | Swin-Unet［58］ | Use Swin Transformer module instead of U-Net's 2D convolutional block | Multi-scale feature fusion， parameters and computational efficiency are optimized for applicability | Lack of extraction of local feature information |
| C2Former［59］ | Designing cross-convolutional self-attention mechanism algorithms to improve semantic feature understanding | Integration of multi-scale feature information and filtering of interfering information | Loss of feature information |
| U-Net+Transformer | Single-scale serial | TransUNet［62］ | Local features are extracted using a CNN and the Transformer module is strung after the CNN to extract global contextual information | Perceiving global information for direct processing of sequential data， adaptation to multi-scale tasks | High computational complexity |
| MultiIB-TransUNet［67］ | A single Transformer layer is used to extract global features to reduce the number of parameters， and multiple IB blocks are utilized to compress noise and improve robustness | Reduced amount of model parameters | Compressing the relevant features， the segmentation accuracy is relatively reduced |
| Multi-scale serial | CoTr［70］ | Transformer in the encoder receives the feature maps at multiple scales in the CNN and introduces a deformable self-attention mechanism that focuses on some of the key sampling points | Computational and spatial complexity is reduced； multi-scale processing of 3D feature maps is realized | Insufficient generalizability |
| HTUNet［73］ | 2D U-Net design MSCAT module is used to jump connections and extract intra-frame features， and the 3D Transformer UNet extracts inter-frame features | Intra- and inter-frame feature fusion for improved segmentation performance | Failure to delineate the nodal region from the gland is computationally expensive |
| Alternating serial | HCTNet［76］ | TEBlocks module is designed to learn global context information and combine it with CNN blocks to extract features | Combining CNN inductive bias for spatial association modeling and Transformer capabilities for remote dependency modeling | Boundary detail segmentation is not effective |
| U-Net+Transformer | Alternating serial | UTNet［79］ | Replace the last convolution in each resolution of U-Net with the Transformer module | Combines convolutional and self-attentive mechanisms that are easy to understand and use | Poor segmentation performance in large-scale training tasks |
| Global parallelism | Transfuse［82］ | CNN and Transformer extract features in parallel， and BiFusion module is designed to fuse features from both branches | Capturing global information while maintaining sensitivity to low-level context， with strong fusion representation | Transformer layer is less efficient and insufficiently generalized |
| HSNet［86］ | Encoder and decoder are connected through an interactive attention mechanism， and the decoder has a two-branch structure | Discriminative remote dependencies can be generated， detail features can be recovered， and generalization ability is high | High model complexity |
| Hierarchical parallelism | UconvTrans［88］ | Each level uses a two-branch structure and a feature fusion module is designed to pass the fused features to the next level | Number of model parameters and the amount of computation are less， which better balances the segmentation accuracy and efficiency | Not applicable to 3D medical image segmentation where slice information is richer |
| RMTF-Net［93］ | Encoder combines Mix Transfomer and RCNN structures， and the feature decoder is designed with a GFI module to re-fuse the feature information extracted by the encoder | Network boundary coding is more capable and a local global balanced feature is available at each level of the encoder | Not extended to 3D image segmentation |
![U-Net structure diagram[12]](/richHtml/lop/2025/62/2/0200001/img_02.jpg)
![Transformer structure diagram[55]](/Images/icon/loading.gif){kind=icon}
