TRiST: Topology-Reinforced Swin Transformer for Medical Image Analysis
Organizations: Department of Computer Science, The University of Texas Rio Grande Valley · School of Mathematical and Statistical Sciences, The University of Texas Rio Grande Valley · Department of Computer Science and Engineering, University of Notre Dame · School of Computing and Informatics, University of Louisiana at Lafayette
Abstract
Medical images contain topological patterns, such as connected components and loops, at scales ranging from small lesions to organ boundaries. Although hierarchical Vision Transformers have advanced medical image analysis, they do not explicitly incorporate topology throughout their multiscale representations. We propose Topology-Reinforced Swin Transformer (TRiST), a framework that integrates topological information into feature construction, refinement, and attention. First, a Hierarchical Multi-Scale Topology Representation algorithm computes 2D persistent homology (H0 and H1) independently at four native Swin stages, producing descriptors aligned with each stage's token grid rather than downsampling fine-scale topology. Second, a Band-Aware H0/H1 Residual Refinement module uses gated residual networks to process short-, medium-, and long-persistence information separately. Third, a Persistence-Band Progressive Attention Bias converts stage-specific descriptors into learned key-saliency scores that directly modify shifted-window attention logits. We instantiate TRiST as Topo-SwinV2-B for classification and Topo-Swin-UNet for segmentation, preserving the underlying Swin-based backbones. Experiments on five medical image datasets show consistent improvements over Swin-based baselines and competitive performance relative to existing topology-augmented methods.
Figures & tables
| ISIC2018 [ 12 ] | CBIS-DDSM [ 30 ] | RetinaMNIST [ 48 ] | ||||
| Method | Acc | AUC | Acc | AUC | Acc | AUC |
| ResNet152 [ 20 ] | 89.63 | 97.43 | 72.29 | 78.74 | 66.78 | 84.97 |
| SENet154 [ 22 ] | 90.52 | 97.85 | 73.14 | 79.40 | 67.44 | 86.45 |
| SwinV2-B [ 33 ] | 90.82 | 98.21 | 72.78 | 79.84 | 67.67 | 86.42 |
| PHG-Net [ 41 ] | 91.60 | 98.44 | 74.57 | 80.93 | 69.00 | 87.18 |
| PHG-Net + multi-scale [ 14 ] | 92.32 | 98.28 | 75.57 | 82.34 | 70.00 | 87.38 |
| Synapse [ 28 ] | ACDC [ 5 ] | |||
| Method | Dice | HD95 | Dice | HD95 |
| U-Net [ 43 ] | 76.05 | 19.02 | 88.32 | 3.58 |
| TransUNet [ 9 ] | 77.34 | 16.83 | 89.42 | 1.58 |
| Attention U-Net [ 40 ] | 77.69 | 17.29 | 86.82 | 5.52 |
| Swin-UNet [ 7 ] | 79.76 | 14.88 | 90.92 | 1.15 |
| V-Net [ 38 ] | 69.01 | 33.69 | 84.74 | 5.15 |
| CBIS-DDSM | Synapse | |||
| Method | Acc | AUC | Dice | HD95 |
| SwinV2-B / Swin-UNet | 72.78 | 79.84 | 79.76 | 14.88 |
| Topo-SwinV2-B w/ PI / Topo-Swin-UNet w/ PI | 75.62 | 82.01 | 82.19 | 11.10 |
| Main Components Ablation | ||||
| W/o Band-Aware H0/H1 Residual Refinement | 73.58 | 80.28 | 81.03 | 11.45 |
| Fusion w/ concatenation | 72.73 | 78.99 | 78.04 | 16.12 |
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
| Stage | Token grid | Stride | Region | Topology representation | Band |
|---|---|---|---|---|---|
| 1 | Short | ||||
| 2 | Medium | ||||
| 3 | Long | ||||
| 4 | Long |
| Module | Parameters | Details |
|---|---|---|
| LayerNorm (98-dim) | 784 | |
| BandAwareH0H1ResidualRefinement | 26,328 | 6 MLPs (hidden dim. 32) + 6 gates + 2 row gates per stage |
| PersistenceBandDecomposer | 0 | fixed index slicing, no learnable parameters |
| PersistenceBandWindowAttentionBias | 12,224 | scoring MLP + per stage |
| TopologyBypassSummary | 43,968 | + + |
| Head expansion ( input dim) | , with |
Medical Image Analysis
Medical image analysis uses computational methods to interpret clinical images, measure anatomy and disease, and support diagnosis and treatment.
See all 309 papers