Research map / Label-efficient and robust learning / Self-supervised and contrastive
Unsupervised Foundation Model-Agnostic Slide-Level Representation Learning
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 COBRA, an unsupervised, image-only contrastive slide encoder that pretrains a Mamba-2 and gated-attention aggregator on frozen tile embeddings from multiple pathology foundation models and magnifications, trained on 3048 TCGA WSIs.
Problem
Learning task-agnostic slide- or patient-level representations of gigapixel pathology WSIs without annotations, since MIL representations are task-specific and SSL augmentations in image space are limited.
Approach
Uses tile embeddings from multiple foundation models and magnifications as feature-space augmentations, and contrastively aligns patient-level embeddings with a Mamba-2 encoder and multi-head gated attention pooling (MoCo-v3 style), enabling inference with any FM including unseen ones.
Key results
- COBRA exceeds state-of-the-art slide encoders on four CPTAC cohorts by at least +4.4% average AUC over PRISM (75.3 vs 70.9 macro-AUC) despite pretraining on only 3048 TCGA WSIs.
- COBRA-V2 (single-FM inference with Virchow2) reaches 75.3% average AUROC, +1.5% over mean Virchow2 patch embeddings.
- Applying COBRA to CTransPath lifts it to 66.4% average AUC, and to GigaPath improves its mean baseline by +2.5% average AUC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use COBRA slide representations instead of PRISM | CPTAC downstream classification · CPTAC (15 downstream tasks, average) | AUC: +4.4% AUC | Helped |
| Use COBRA slide representations instead of Virchow2 mean patch embeddings | CPTAC downstream classification · CPTAC (15 downstream tasks, average) | AUC: +1.5% | Helped |
| Pretrain using three magnifications instead of only 0.5 MPP (ablation) | COBRA · CPTAC downstream tasks (average across models) | AUC: +1.73% AUC | Helped |
| Use 5× inference instead of 20× inference (ablation) | COBRA†-V2 · CPTAC downstream tasks (average) | AUC: +3.8% | Helped |
| Use 9× inference instead of 20× inference (ablation) | COBRA†-V2 · CPTAC downstream tasks (average) | AUC: +3.7% | Helped |
| Combine embeddings from all pretrained FMs at inference instead of using a single FM (ablation) | COBRA · CPTAC downstream tasks at 2 MPP (average) | AUC: +0.53% AUC | Helped |
| Use unseen GigaPath patch embeddings with COBRA rather than mean patch embeddings (ablation) | COBRA · CPTAC downstream tasks (average) | AUC: +2.5% average AUC | Helped |
Surprising findings
- Gigapath-SE slide encoder underperforms its own GigaPath mean patch embedding baseline, unlike other slide encoders.
- Using multiple magnifications during pretraining gives only modest average gains (+1.73% AUC) and COBRA trained on 0.5 MPP alone does not always beat the multi-magnification version.
Limitations
- Pretraining limited to 3048 TCGA WSIs from a few cancer types; generalization to other tissues is untested beyond the evaluated cohorts.
- Only contrastive SSL objective explored; authors note other SSL objectives are future work.
- Several CPTAC tasks show high variance across folds (e.g., KRAS, PIK3CA).
Experimental setup
Datasets: TCGA-BRCA TCGA-CRC TCGA-LUAD TCGA-LUSC TCGA-STAD CPTAC-BRCA CPTAC-COAD CPTAC-LUAD CPTAC-LUSC
Benchmarks: CPTAC-BRCA CPTAC-COAD CPTAC-LUAD CPTAC-LUSC
Metrics: AUROC AUPRC F1 balanced accuracy
Baselines: PRISM GigaPath-SE MADELEINE CHIEF CTransPath Virchow2 UNI H-Optimus-0 CONCH Virchow
Builds on: MoCo-v3 Mamba-2 MambaMIL ABMIL InfoNCE
Closest papers by meaning
- Transcriptomics-guided Slide Representation Learning in Computational Pathology (CVPR 2024 · similarity 0.90)
- PathVQ: Reforming Computational Pathology Foundation Model for Whole Slide Image Analysis via Vector Quantization (NeurIPS 2025 · similarity 0.90)
- Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology (CVPR 2024 · similarity 0.89)
- Rotation-Agnostic Image Representation Learning for Digital Pathology (CVPR 2024 · similarity 0.88)
- Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis (CVPR 2024 · similarity 0.87)
- Nonparametric Distribution Matching for Self-Supervised Whole-Slide Image Condensation (NeurIPS 2026 · 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.