Research map / Architectures and efficiency / Attention and transformers
Bridging Local Inductive Bias and Long-Range Dependencies with Pixel-Mamba for End-to-end Whole Slide Image Analysis
Zhongwei Qiu, Hanqing Chao, Tiancheng Lin, Wanxing Chang, Zijiang Yang, Wenpei Jiao, Yixuan Shen, Yunshuo Zhang, Yelin Yang, Wenbin Liu, Hui Jiang, Yun Bian and 3 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add hierarchical token scales 1-32 instead of fixed 16+32 tokens (ablation) | Vision-Mamba-Surv · TCGA-BLCA survival | C-index: 0.6116 -> 0.6507 | Improved |
| combine token scales 16 and 32 (ablation) | Vision-Mamba-Surv · TCGA-BLCA survival | C-index: 0.5977 -> 0.6116 | Improved |
| use Mamba long-sequence model instead of LongViT dilated attention (ablation) | end-to-end WSI survival · TCGA-BLCA survival | C-index: 0.5743 -> 0.5977 | Improved |
| add hierarchical slide representation (Pixel-Mamba) to Vision Mamba (ablation) | Vision-Mamba-Surv · TCGA-BLCA survival | C-index: 0.5977 -> 0.6507 | Improved |
| use scan window 224 instead of 32 (ablation) | Pixel-Mamba-Surv · TCGA-BLCA survival | C-index: 0.6083 -> 0.6507 | Improved |
| enlarge scan window from 224 to 320 (ablation) | Pixel-Mamba-Surv · TCGA-BLCA survival | C-index: 0.6507 -> 0.6242 | Hurt |
| raise magnification from 0.6x to 2.5x (ablation) | Pixel-Mamba-Surv · TCGA-BLCA survival | C-index: 0.5855 -> 0.6507 | Improved |
| raise magnification from 1.0x to 2.5x (ablation) | Pixel-Mamba-Surv · TCGA-BLCA survival | C-index: 0.6307 -> 0.6507 | Improved |
| use pathology foundation model features (GigaPath) instead of ImageNet ResNet-50 | ABMIL two-stage · TCGA-BLCA tumor staging | macro F1: 0.4100 -> 0.5220 | Improved |
| use end-to-end Pixel-Mamba (ImageNet only) instead of best two-stage GigaPath ILRA-MIL | Pixel-Mamba-Stage · TCGA-BLCA tumor staging | macro F1: 0.5320 -> 0.5334 | Improved |
| increase LongViT input scale from 0.6x to 2.5x | LongViT · TCGA-BLCA tumor staging | macro F1: 0.2310 -> 0.4963 | Improved |
| use Pixel-Mamba-6M instead of Vim-Ti | Vim-Ti · ImageNet-1K val | top-1 accuracy (%): 76.1 -> 77.8 | Improved |
Closest papers by meaning
- Rethinking Transformer for Long Contextual Histopathology Whole Slide Image Analysis (NeurIPS 2024 · similarity 0.87)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.87)
- 2DMamba: Efficient State Space Model for Image Representation with Applications on Giga-Pixel Whole Slide Image Classification (CVPR 2025 · similarity 0.86)
- S³-Mamba: Small-Size-Sensitive Mamba for Lesion Segmentation (AAAI 2025 · similarity 0.86)
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.85)
- M3amba: Memory Mamba is All You Need for Whole Slide Image Classification (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
