Research map / Label-efficient and robust learning / Self-supervised and contrastive
Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation
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.
Reformulates self-supervised WSI condensation as distribution matching under a frozen encoder and proposes NICER, a nonparametric (PPP-MAP-motivated) two-stage prototype condensation method with slide-adaptive capacity.
Problem
Existing self-supervised WSI condensation methods rely on heuristic prototype learning and do not explicitly preserve the encoder-induced feature distribution needed by downstream MIL predictors, while WSIs are too large (thousands to millions of patches) for cheap training and storage.
Approach
Casts WSI condensation as optimal-transport distribution matching between patch embeddings from a fixed pretrained encoder and condensed prototypes, shows the per-slide objective equals a MAP problem under a Poisson point process prior with cardinality penalty, and approximates it via alternating pattern exploration and concept condensation with active-set pruning of unused concepts.
Key results
- NICER achieves an average accuracy improvement of 7.44% over prior methods across five histopathology datasets (abstract).
- On PANDA with ABMIL, NICER reaches Kappa 90.98 vs 65.52 for PANTHER, and on NSCLC with ILRA it reaches F1 94.23 vs 90.69 for PANTHER.
- On NSCLC with ILRA, NICER compresses storage from 194 GB to 5.7 GB versus 3.2 GB for PANTHER, at higher condensation cost (1.51 s vs 0.05 s).
Surprising findings
- With the ILRA predictor, NICER surpasses the Whole-Bag upper bound on several survival tasks.
- NICER's condensation cost per WSI is higher than PANTHER (1.51 s vs 0.05 s), trading compute for accuracy.
Limitations
- Bag-of-features formulation ignores spatial and multi-scale dependencies among patches.
- PPP prior assumes independent concept occurrence, which may not capture correlated morphological patterns.
- Evaluated only with fixed pretrained encoders (two foundation models).
Experimental setup
Datasets: PANDA TCGA NSCLC TCGA-BRCA TCGA-LUAD CPTAC
Benchmarks: PANDA NSCLC TCGA-BRCA TCGA-LUAD CPTAC
Metrics: Cohen's Kappa accuracy weighted F1 balanced accuracy C-Index
Baselines: DeepSets ProtoCount H2T InfiniteGPFA OT PANTHER ABMIL DSMIL ILRA
Builds on: PANTHER optimal transport UNI ABMIL DSMIL
Closest papers by meaning
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.88)
- HistoFS: Non-IID Histopathologic Whole Slide Image Classification via Federated Style Transfer with RoI-Preserving (CVPR 2025 · similarity 0.88)
- Rethinking Transformer for Long Contextual Histopathology Whole Slide Image Analysis (NeurIPS 2024 · similarity 0.88)
- Sequential Attention-based Sampling for Histopathological Analysis (NeurIPS 2025 · similarity 0.88)
- Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis (CVPR 2024 · similarity 0.88)
- Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning (CVPR 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.