Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
FAST: A Dual-tier Few-Shot Learning Paradigm for Whole Slide Image Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FAST, a dual-tier few-shot WSI classification paradigm that annotates few slides and few patches, combining a learnable key-value cache branch with a CoOp-style CLIP text prompt branch.
Problem
Few-shot whole slide image (WSI) classification where fine-grained patch annotation is expensive and few WSIs are available; existing few-shot methods use only sparse patch labels or weak slide labels and underuse unlabeled patches.
Approach
A dual-level annotation scheme (few slides, few labeled patches per slide) plus a cache branch whose keys and values (pseudo-labels of unlabeled patches) are both learnable, fused with a prompt-learned CLIP text-classifier branch; instance predictions are pooled to bag level.
Key results
- On CAMELYON16 with 16 bag shots and 16 instance shots, FAST reaches 0.9151 instance-level and 0.8197 bag-level AUC versus 0.7227 and 0.5965 for Tip-Adapter-F, against 0.9532 and 0.8555 for full supervision.
- FAST uses about 0.22% of the patch annotation cost of full supervision on CAMELYON16 at 16 bag shots.
- On TCGA-RENAL at 16 bag shots, FAST's mean bag-level AUC is 0.9235 versus 0.9438 for full supervision.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FAST instead of Tip-Adapter-F | CAMELYON16, 16 bag shots / 16 instance shots · CAMELYON16, 16 bag shots / 16 instance shots | instance-level AUC: 0.7227±0.0098 -> 0.9151±0.0200 | Helped |
| use FAST instead of Tip-Adapter-F | CAMELYON16, 16 bag shots / 16 instance shots · CAMELYON16, 16 bag shots / 16 instance shots | bag-level AUC: 0.5965±0.0243 -> 0.8197±0.0474 | Helped |
| use FAST instead of Tip-Adapter-F | TCGA-RENAL, 16 bag shots / 16 instance shots · TCGA-RENAL, 16 bag shots / 16 instance shots | mean bag-level AUC: 0.8459 -> 0.9235 | Helped |
| make the cache feature learnable (ablation) | FAST cache branch · CAMELYON16, 16 bag shots / 16 instance shots | instance-level AUC: 0.6835±0.0135 -> 0.7227±0.0098 | Helped |
| make the cache feature learnable (ablation) | FAST cache branch · CAMELYON16, 16 bag shots / 16 instance shots | bag-level AUC: 0.4913±0.0164 -> 0.5965±0.0243 | Helped |
| add a learnable label cache using unlabeled instances (ablation) | FAST cache branch · CAMELYON16, 16 bag shots / 16 instance shots | instance-level AUC: 0.7227±0.0098 -> 0.9165±0.0213 | Helped |
| add a learnable label cache using unlabeled instances (ablation) | FAST cache branch · CAMELYON16, 16 bag shots / 16 instance shots | bag-level AUC: 0.5965±0.0243 -> 0.8183±0.0560 | Helped |
| add the prior branch to the cache branch (ablation) | FAST · CAMELYON16, 16 bag shots / 16 instance shots | instance-level AUC: 0.9165±0.0213 -> 0.9151±0.0200 | No clear effect |
| add the prior branch to the cache branch (ablation) | FAST · CAMELYON16, 16 bag shots / 16 instance shots | bag-level AUC: 0.8183±0.0560 -> 0.8197±0.0474 | No clear effect |
| vary the core set size from 100 to 5000 (ablation) | FAST · CAMELYON16, 4, 8, and 16 bag shots | AUC | Mixed |
| increase annotated instances per bag from 4 or 16 to 64 (ablation) | FAST · CAMELYON16, 1, 2, or 4 bags per class | AUC | Helped |
| increase annotated instances per bag from 4 or 16 to 64 (ablation) | FAST · CAMELYON16, 8 or 16 bags per class | AUC | No clear effect |
Surprising findings
- Zero-shot CONCH (a pathology VLM) outperforms FAST on CAMELYON16 instance-level AUC at 0.8929 versus FAST's few-shot results at low bag counts, and on TCGA-RENAL zero-shot CONCH's mean bag AUC (0.9141) exceeds FAST at 1 to 8 bag shots.
- Using the prior branch alone outperforms the cache branch alone at 1-2 bag shots, while the cache branch dominates at 4 or more bags.
Limitations
- Bag-level prediction is a simple pooling of instance predictions, ignoring relations between instances.
- Evaluated only on two datasets (CAMELYON16 binary and TCGA-RENAL three-class).
- Large variance across random samplings at low bag counts (e.g., bag-level AUC std up to 0.0846 at 1 bag shot).
Experimental setup
Datasets: CAMELYON16 TCGA-RENAL
Benchmarks: CAMELYON16 TCGA-RENAL
Metrics: instance-level AUC bag-level AUC
Baselines: Tip-Adapter Tip-Adapter-F Zero-shot CLIP Zero-shot PLIP Zero-shot CONCH Fully supervised
Builds on: Tip-Adapter CoOp CLIP
Closest papers by meaning
- Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction (CVPR 2024 · similarity 0.88)
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.88)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.87)
- CaMIL: Causal Multiple Instance Learning for Whole Slide Image Classification (AAAI 2024 · similarity 0.87)
- ViLa-MIL: Dual-scale Vision-Language Multiple Instance Learning for Whole Slide Image Classification (CVPR 2024 · similarity 0.87)
- OODML: Whole Slide Image Classification Meets Online Pseudo-Supervision and Dynamic Mutual Learning (AAAI 2025 · similarity 0.87)
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.