Research map / Vision-language and multimodal / CLIP and zero-shot
MAPLE: Multi-scale Attribute-enhanced Prompt Learning 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 MAPLE, a hierarchical LLM-prompted framework that combines multi-scale entity-level and slide-level prompt alignment with cross-scale entity graph learning for few-shot WSI classification.
Problem
Few-shot WSI classification with VLM prompt learning relies on slide-level prompts and misses subtype-specific phenotypic attributes of histological entities (nuclei, glands) across magnification scales.
Approach
LLM-generated multi-scale entity prompts with subtype-specific attributes drive entity-level cross-attention features and cross-scale GAT entity graphs; entity-level and slide-level logits are fused for prediction.
Key results
- 16-shot TCGA-NSCLC: MAPLE AUC 90.3%, F1 80.6%, ACC 81.0% with PLIP, versus FOCUS 86.2% AUC.
- 16-shot TCGA-BRCA: MAPLE AUC 80.1% vs ViLa-MIL 78.9%; 16-shot TCGA-RCC AUC 96.9% vs FOCUS 95.1%.
- Ablations on 16-shot show each of instance selection, entity cross-attention, and entity graph contributes; entity-only and slide-only variants are both below the full model.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MAPLE instead of the strongest listed baseline | few-shot WSI classification with PLIP · TCGA-BRCA 4-, 8-, and 16-shot; TCGA-RCC 4-, 8-, and 16-shot; TCGA-NSCLC 4-, 8-, and 16-shot | AUC: TCGA-BRCA: MAPLE 0.722 ± 0.063 vs FOCUS 0.703 ± 0.051 (4-shot); 0.786 ± 0.070 vs ViLa-MIL 0.770 ± 0.062 (8-shot); 0.801 | Helped |
| Integrate low- and high-resolution entities rather than using one scale only (ablation) | MAPLE · TCGA-BRCA 16-shot | AUC: 0.790 ± 0.052 (MAPLE-Low); 0.779 ± 0.046 (MAPLE-High); 0.801 ± 0.031 (MAPLE) | Helped |
| Integrate low- and high-resolution entities rather than using one scale only (ablation) | MAPLE · TCGA-RCC 16-shot | AUC: 0.949 ± 0.009 (MAPLE-Low); 0.956 ± 0.016 (MAPLE-High); 0.969 ± 0.014 (MAPLE) | Helped |
| Integrate low- and high-resolution entities rather than using one scale only (ablation) | MAPLE · TCGA-NSCLC 16-shot | AUC: 0.866 ± 0.045 (MAPLE-Low); 0.875 ± 0.050 (MAPLE-High); 0.903 ± 0.033 (MAPLE) | Helped |
| Combine entity-level and slide-level logits (ablation) | MAPLE · TCGA-BRCA 16-shot | AUC: 0.792 ± 0.052 (MAPLE-Entity); 0.789 ± 0.036 (MAPLE-Slide); 0.801 ± 0.031 (MAPLE) | Helped |
| Combine entity-level and slide-level logits (ablation) | MAPLE · TCGA-RCC 16-shot | AUC: 0.960 ± 0.014 (MAPLE-Entity); 0.951 ± 0.011 (MAPLE-Slide); 0.969 ± 0.014 (MAPLE) | Helped |
| Combine entity-level and slide-level logits (ablation) | MAPLE · TCGA-NSCLC 16-shot | AUC: 0.879 ± 0.060 (MAPLE-Entity); 0.891 ± 0.033 (MAPLE-Slide); 0.903 ± 0.033 (MAPLE) | Helped |
| Add language-guided instance selection (ablation) | MAPLE · TCGA-BRCA 16-shot | AUC: 0.792 ± 0.059 -> 0.801 ± 0.031 | Helped |
| Add language-guided instance selection (ablation) | MAPLE · TCGA-RCC 16-shot | AUC: 0.962 ± 0.015 -> 0.969 ± 0.014 | Helped |
| Add language-guided instance selection (ablation) | MAPLE · TCGA-NSCLC 16-shot | AUC: 0.885 ± 0.064 -> 0.903 ± 0.033 | Helped |
| Add entity-guided cross-attention (ablation) | MAPLE · TCGA-BRCA 16-shot | AUC: 0.784 ± 0.019 -> 0.801 ± 0.031 | Helped |
| Add entity-guided cross-attention (ablation) | MAPLE · TCGA-RCC 16-shot | AUC: 0.959 ± 0.015 -> 0.969 ± 0.014 | Helped |
Limitations
- Relies on GPT-4 generated prompts whose quality is not systematically validated.
- Results on few-shot splits show large standard deviations, and gains over strongest baselines are often within variance.
- Code is promised only upon acceptance, with no released artifacts.
Experimental setup
Datasets: TCGA-BRCA TCGA-RCC TCGA-NSCLC
Benchmarks: TCGA-BRCA TCGA-RCC TCGA-NSCLC
Metrics: AUC F1 ACC
Baselines: ABMIL TransMIL GTMIL WiKG TOP ViLa-MIL MSCPT FOCUS
Builds on: CLIP CoOp TOP ViLa-MIL CLAM
Closest papers by meaning
- FOCUS: Knowledge-enhanced Adaptive Visual Compression for Few-shot Whole Slide Image Classification (CVPR 2025 · similarity 0.90)
- Few-Shot Learning from Gigapixel Images via Hierarchical Vision-Language Alignment and Modeling (NeurIPS 2025 · similarity 0.88)
- Memory-Efficient Prompt Tuning for Incremental Histopathology Classification (AAAI 2024 · similarity 0.87)
- Generalizable Whole Slide Image Classification with Fine-Grained Visual-Semantic Interaction (CVPR 2024 · similarity 0.87)
- Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis (CVPR 2024 · similarity 0.86)
- BiomedCoOp: Learning to Prompt for Biomedical Vision-Language Models (CVPR 2025 · similarity 0.86)
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.