Research map / Architectures and efficiency / Attention and transformers
Turning Pre-Trained Vision Transformers into End-to-End Histopathology Whole Slide Image Models for Survival Prediction
Jiawen Li, Jiali Hu, Xitong Ling, Renao Yan, Yuxuan Chen, Tian Guan, Yonghong He
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 |
|---|---|---|---|
| convert pre-trained ViT to end-to-end WSI model (E2E-ViT) instead of two-stage MIL | H0-mini · 5-dataset survival overall | C-index: 0.6810 -> 0.7158 | Improved |
| convert pre-trained ViT to E2E-ViT instead of best two-stage MIL | CONCH · 5-dataset survival overall | C-index: 0.6902 -> 0.6978 | Improved |
| convert ImageNet ViT-Small to E2E-ViT instead of two-stage MIL | ViT-Small · 5-dataset survival overall | C-index: 0.6386 -> 0.6667 | Improved |
| use E2E-ViT linear-probe slide features instead of slide foundation models | E2E-ViT · 5-dataset survival overall (linear probe) | C-index: H0-mini 0.6685, CONCH 0.6534, ViT-Small 0.5959 vs TITAN 0.6582 | Mixed |
| use E2E H0-mini instead of SFMs for slide encoding speed | E2E H0-mini · Inference time, 10,000 tiles | speedup: 2.64x faster than CHIEF, 7.49x faster than FEATHER | Improved |
| replace mean patch merging with max pooling (ablation) | E2E-ViT · HNSC/LUAD/PDAC survival | C-index: CONCH HNSC 0.7421 -> 0.7387 | Hurt |
| replace mean patch merging with attention pooling (ablation) | E2E-ViT · HNSC/LUAD/PDAC survival | C-index: CONCH HNSC 0.7421 -> 0.7321 | No clear effect |
| use ALiBi relative positional encoding instead of none (ablation) | E2E CONCH · LUAD survival | C-index: 0.6350 -> 0.7077 | Improved |
| use ALiBi instead of learnable positional embedding (ablation) | E2E-ViT · HNSC/LUAD/PDAC survival, 3 backbones | C-index: CONCH HNSC 0.6920 -> 0.7421; ViT-Small PDAC 0.6710 -> 0.6408 | Mixed |
| vary ratio of sequence length to tile number (ablation) | E2E-ViT · HNSC, LUAD, PDAC, CCRCC | C-index | No clear effect |
| convert larger pathology ViTs (UNI, Prov-GigaPath, Virchow, PathOrchestra, UNI-2) instead of two-stage ABMIL | E2E-ViT · LUAD survival (fine-tuning) | C-index: +1.03% to +3.87% | Improved |
| convert larger ViTs under linear probing, versus TITAN | E2E-ViT · MBC survival (linear probe) | C-index: TITAN 0.6624; Virchow, PathOrchestra, UNI-2 surpass it | Mixed |
Closest papers by meaning
- Transformer-Based Video-Structure Multi-Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.90)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.90)
- Rethinking Transformer for Long Contextual Histopathology Whole Slide Image Analysis (NeurIPS 2024 · similarity 0.90)
- Interpretable Vision-Language Survival Analysis with Ordinal Inductive Bias for Computational Pathology (ICLR 2025 · similarity 0.89)
- Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning (CVPR 2025 · similarity 0.88)
- Free Lunch in Pathology Foundation Model: Task-specific Model Adaptation with Concept-Guided Feature Enhancement (NeurIPS 2024 · similarity 0.88)
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.
