TransClaw U-Net: Claw U-Net with Transformers for Medical Image Segmentation

• CNN-based methods and U-Net variants have achieved strong medical-image segmentation results, and accurate segmentation is important for clinical diagnosis [1–12].The introduction names organ, lesion, and cardiac segmentation among the application areas.Source: §1 · p. 1
• CNN operations focus on local features and may miss long-range dependencies, limiting accurate segmentation in some image regions [13].The introduction also says cited alternatives such as self-attention, atrous convolution, and image pyramids do not address long-range dependencies in medical images.Source: §1 · p. 1
• Transformers model long-term sequence features, and prior work has applied them to vision and medical segmentation, including within U-Net encoders [13] [19] [21] [22].The cited approaches include patch encoding and a pure-transformer U-Net-like architecture.Source: §1 · pp. 1–2
• TransClaw U-Net combines Claw U-Net’s detail segmentation with transformer feature learning; its Synapse experiments report more accurate organ segmentation, while larger inputs, smaller patches, and more skip-connections require more computing resources.The design uses convolutional feature maps to encode patches, transformer self-attention for global connections, and retained bottom upsampling in the decoder.Source: §1 · p. 2

• CNNs and U-shaped networks form a major medical-segmentation backbone, with variants using architectural and convolutional modifications [5, 6, 10, 23–25].Examples include Resnet34-UNet, Attention-UNet, Dense-UNet, UNet++, deformable convolution, and deep convolution.Source: §2 · p. 2
• Prior work combines self-attention or transformers with CNNs to address long-range dependence, including inserting transformers into U-Net structures [13].The paper adopts Claw U-Net [4] as its framework and introduces a transformer to examine its effect on medical-image segmentation.Source: §2 · p. 2
• Transformers model long-term sequence features and have been applied to image tasks, but the text notes that transformer structures require substantial pretraining data compared with trainable U-Net structures on small datasets [20] [21] [28].The proposed design uses Claw U-Net as its skeleton, adds transformer encoding to convolutional feature maps, and combines encoded and upsampling features in the decoder.Source: §2 · p. 2

3. METHOD

• Self-attention in the encoding path targets global feature learning, while the bottom-upsampling and decoding path maintains detailed information for segmentation.Source: §3 · p. 2

3.1. Overall structure of TransClaw U-Net

• The encoder, bottom upsampling, and decoder are linked by skip connections; upsampling restores the image to its initial size.Source: §3.1 · pp. 2–3

3.2. Transformer in TransClaw U-Net

• Patch vectors form an N × P² sequence for transformer blocks; dividing attention dimensions keeps computational cost not much different from single-head attention.Source: §3.2 · p. 4

4.1. Datasets

• The Synapse evaluation uses 30 abdominal CT cases, with 18 cases for training and 12 for testing, and assesses eight organs using DSC and HD.The dataset contains 3,779 axial images.Source: §4.1 · p. 4

4.2. Implementation Details

• Experiments use 224 × 224 inputs, batch size 24, and 150 training epochs, with SGD and cross-entropy loss; the best validation-set model is selected.All experiments use the same parameter settings for fair comparison.Source: §4.2 · p. 4

4.3. Comparison with other models on Synapse dataset

• TransClaw U-Net leads the Synapse comparison with DSC 78.09% and HD 26.38%, improving over TransUNet by 0.6% DSC and 5.3% HD and leading on three of eight organs.The other compared networks have DSC below 77.5% and HD above 31.5.Source: §4.3 · p. 5
• TransClaw U-Net shows better organ-boundary segmentation than TransUNet, especially between small-area organs; Attention-UNet over-segments and U-Net is inaccurate.The authors attribute TransClaw U-Net’s performance to Claw U-Net’s detail preservation and the transformer’s global-context learning.Source: §4.3 · p. 5, Figure 2 · p. 5

4.4. Ablation Experiments

• Increasing input resolution from 224 × 224 to 512 × 512 significantly raises DSC, but requires more computation and training time; the authors adopt 224 × 224.At unchanged patch size, larger inputs also lengthen the transformer sequence.Source: §4.4 · pp. 5–6
• Reducing skip-connections worsens segmentation; removing them drops DSC by 10.8% for Arota and 10.16% for Pancreas.The authors describe skip-connections as combining encoding, bottom upsampling, and decoding parts.Source: §4.4 · p. 6
• With 16×16 patches, DSC is 78.09%, compared with 77.27% for 24×24 patches; larger patches shorten the transformer sequence, while smaller patches make learned parameters grow rapidly and occupy computing resources.Source: §4.4 · p. 6

5. CONCLUSION

• TransClaw U-Net achieves excellent performance on the Synapse multi-organ CT dataset, while ablation experiments show excellent generalization performance.The conclusion reports these outcomes without giving metric values here.Source: §5 · p. 6
• The encoder combines CNN convolution with transformer self-attention to obtain global context and address convolution’s limitations in capturing it.The decoder retains the upsampling structure to improve medical-image segmentation accuracy.Source: §5 · p. 6
• Fig. 3 compares performance across different numbers of skip-connections.Source: Figure 3 · p. 7