Table 1 Summary of recent attention-based histopathology models compared to the proposed HistoDARE.
From: A histopathology aware DINO model with attention based representation enhancement
Model | Backbone | Attention type | Key remarks/limitations |
|---|---|---|---|
HistoSSL18 | Multi-branch SSL | Multi-level fusion | Reduces annotation need; lacks spatial explainability |
MMAE19 | MAE + H&E/RGB | Dual-modality | Enhances morphology; requires stain alignment |
CycleGAN-SSL20 | GAN-based SSL | Cycle-consistent | Improves domain transfer; high training cost |
TransFuse21 | CNN + Transformer | Parallel fusion | Strong accuracy; dual-branch inference overhead |
DS-TransUNet22 | Swin Transformer | Dual-scope | Good context modeling; high memory need |
EG-TransUNet23 | Swin + Attention | Enhancement module | Improved mDice; limited validation |
FCB-SwinV224 | SwinV2 hybrid | Channel fusion | Very deep; heavy computation |
SAG-ViT25 | EfficientNet + GAT | Graph attention | High F1; multi-stage and costly |
SAM/ViT | Prompt-based | Zero-shot; requires user interaction | |
Bilal & Asif (2025)17 | CNN + SA | Lightweight fusion | Efficient; dataset-specific |
Hekmat (2025)16 | CNN + Transformer | Sequential attention | More interpretable; higher complexity |
HistoDARE (Ours) | DINOv2 (ViT-L/14) | Hierarchical (Spatial + Channel + Residual) | Unified design; interpretable and efficient |
