Research map / Vision-language and multimodal / Human and embodied modeling
FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot 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 FOCUS, a three-stage adaptive visual token compression MIL framework that combines pathology foundation model features, language prompts from an LLM, and neighbor-aware filtering for few-shot WSI classification.
Problem
Few-shot whole slide image classification in computational pathology, where most WSI patches are non-diagnostic and dilute learning from very few labeled slides, and existing methods underuse pathology foundation models and align language prompts with all patches at once.
Approach
Progressively compress the patch sequence: FM-feature sliding-window redundancy removal, then language-prompt-guided top-k token selection, then pairwise cosine-similarity filtering, followed by text-guided cross-modal aggregation for prediction.
Key results
- On TCGA-NSCLC 4-shot, FOCUS reaches 81.9% Balanced ACC and 91.5% AUC, vs. ViLa-MIL at 80.7% and 88.5%.
- On CAMELYON 4-shot, FOCUS reaches 70.1% Balanced ACC vs. 65.8% for ViLa-MIL.
- On UBC-OCEAN 16-shot, FOCUS reaches 96.7% AUC, 1.1 points above the prior best (95.6%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FOCUS instead of ViLa-MIL | Few-shot WSI classification · TCGA-NSCLC 4-shot | Balanced ACC: 0.807±0.043 -> 0.819±0.044 | Helped |
| Use FOCUS instead of ViLa-MIL | Few-shot WSI classification · CAMELYON 4-shot | Balanced ACC: 0.658±0.036 -> 0.701±0.082 | Helped |
| Use FOCUS instead of Attentive Pooling | Few-shot WSI classification · UBC-OCEAN 4-shot | Balanced ACC: 0.671±0.080 -> 0.704±0.088 | Helped |
| Add a language prompt branch to BaseMIL (ablation) | FOCUS · UBC-OCEAN 4-shot | Balanced ACC: 0.623±0.086 -> 0.638±0.074; 0.682±0.034 -> 0.698±0.049; 0.779±0.040 -> 0.798±0.038 | Helped |
| Add KAVTC (ablation) | FOCUS Prompt variant · UBC-OCEAN | Balanced ACC: 0.638±0.074 -> 0.673±0.082; 0.698±0.049 -> 0.731±0.037; 0.798±0.038 -> 0.827±0.030 | Helped |
| Add SVTC (ablation) | FOCUS Prompt + KAVTC · UBC-OCEAN | Balanced ACC: 0.673±0.082 -> 0.688±0.056; 0.731±0.037 -> 0.769±0.045; 0.827±0.030 -> 0.849±0.048 | Helped |
| Add CrossAgg (ablation) | FOCUS Prompt + KAVTC + SVTC · UBC-OCEAN | Balanced ACC: 0.688±0.056 -> 0.704±0.088; 0.769±0.045 -> 0.773±0.054; 0.849±0.048 -> 0.864±0.032 | Helped |
| Use CONCH instead of other evaluated pathology foundation models (ablation) | FOCUS · UBC-OCEAN 4-, 8-, and 16-shot | Balanced ACC: 70.4%, 77.3%, and 86.4% | Helped |
| Use PLIP as the feature encoder (ablation) | FOCUS · UBC-OCEAN few-shot settings | Balanced ACC | Hurt |
| Generate pathology descriptions with Claude-3.5-Sonnet instead of other evaluated LLMs (ablation) | FOCUS · UBC-OCEAN 16-shot | Balanced ACC: 86.4% | Helped |
| Generate pathology descriptions with ChatGPT3.5-Turbo (ablation) | FOCUS · UBC-OCEAN 16-shot | Balanced ACC: 84.8% | Hurt |
Surprising findings
- PLIP, a vision-language pathology model, yields the lowest performance among the evaluated foundation models on UBC-OCEAN.
- Performance gap between foundation models narrows at 16-shot (UNI, GPFM, and Virchow reach comparable results).
Limitations
- Relies on LLM-generated class descriptions whose quality affects results, with performance varying across LLMs.
- Evaluated on three public datasets only, with results reported for a limited set of shot counts.
Experimental setup
Datasets: TCGA-NSCLC CAMELYON16 CAMELYON17 UBC-OCEAN
Benchmarks: TCGA-NSCLC CAMELYON UBC-OCEAN
Metrics: Balanced ACC AUC F1 score
Baselines: Mean Pooling Max Pooling Attentive Pooling TransMIL DS-MIL DTFD-MIL WiKG-MIL TOP-MIL ViLa-MIL
Builds on: ViLa-MIL TOP CLAM CONCH
Closest papers by meaning
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.92)
- MAPLE: Multi-scale Attribute-enhanced Prompt Learning for Few-shot Whole Slide Image Classification (NeurIPS 2025 · similarity 0.90)
- CPath-Omni: A Unified Multimodal Foundation Model for Patch and Whole Slide Image Analysis in Computational Pathology (CVPR 2025 · similarity 0.89)
- MLLM-HWSI: A Multimodal Large Language Model for Hierarchical Whole Slide Image Understanding (CVPR 2026 · similarity 0.89)
- PathFLIP: Fine-grained Language-Image Pretraining for Versatile Computational Pathology (AAAI 2026 · similarity 0.89)
- Multi-Resolution Pathology-Language Pre-training Model with Text-Guided Visual Representation (CVPR 2025 · similarity 0.89)
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.