Phase 1: Foundation Model Development
Duration: June - October 2025
Objective: Develop self-supervised learning approaches for pathology foundation models
MSK-SLCPFM Dataset
The Phase 1 dataset provides participants with unprecedented access to the MSK-SLCPFM collection - the largest pathology dataset ever assembled for competition purposes.
Dataset Specifications
- Scale: ~300 million pathology image tiles
- Source: 51,578 whole slide images (WSIs)
- Coverage: 39 cancer types
- Origins: TCGA, CPTAC, and MSK repositories
- Scanner Diversity: Multiple vendors (Aperio, Hamamatsu, Leica, Parmana, Phillips)
Data Composition
| Source | Number of WSIs |
|---|---|
| TCGA | 10,438 |
| CPTAC | 6,004 |
| MSK | 34,690 |
| Total | 51,132 |
Multi-Resolution Tile Formats
To capture multi-scale tissue contexts essential for pathological analysis, the dataset includes three distinct tile formats:
- 224 × 224 pixel tiles from 20× WSIs (0.5μm²/pixel)
- 448 × 448 pixel tiles from 40× WSIs (0.25μm²/pixel)
- 1024 × 1024 pixel tiles from 20× WSIs with independent coordinates
This strategic sampling approach facilitates hierarchical tissue structure learning across varying magnifications, mimicking the multi-resolution examination methodology employed by pathologists.
Complete Cancer Types Distribution
The dataset spans all 39 cancer types with the following distribution:
| # | Cancer Type | Abbreviation | TCGA | CPTAC | MSK | Total |
|---|---|---|---|---|---|---|
| 1 | Adrenocortical Carcinoma | ACC | 79 | 0 | 50 | 129 |
| 2 | Anal Squamous Cell Carcinoma | ANSC | 0 | 0 | 50 | 50 |
| 3 | Appendiceal Cancer | APAD | 0 | 0 | 50 | 50 |
| 4 | Bladder Urothelial Carcinoma | BLCA | 412 | 0 | 2,000 | 2,412 |
| 5 | Bone Cancer | BOCA | 0 | 0 | 50 | 50 |
| 6 | Breast Invasive Carcinoma | BRCA | 1,097 | 0 | 3,000 | 4,097 |
| 7 | Cervical Squamous Cell Carcinoma | CESC | 307 | 0 | 200 | 507 |
| 8 | Cholangiocarcinoma | CHOL | 45 | 0 | 1,000 | 1,045 |
| 9 | Colon Adenocarcinoma | COAD | 461 | 0 | 3,000 | 3,461 |
| 10 | Lymphoid Neoplasm Diffuse Large B-cell Lymphoma | DLBC | 58 | 0 | 500 | 558 |
| 11 | Esophageal Carcinoma | ESCA | 185 | 0 | 200 | 385 |
| 12 | Gastrointestinal Stromal Sarcoma | GIST | 0 | 0 | 500 | 500 |
| 13 | Glioblastoma Multiforme | GBM | 528 | 462 | 1,000 | 1,990 |
| 14 | Head and Neck Squamous Cell Carcinoma | HNSC | 478 | 390 | 500 | 1,368 |
| 15 | Kidney Chromophobe | KICH | 113 | 0 | 50 | 163 |
| 16 | Kidney Renal Clear Cell Carcinoma | KIRC | 537 | 783 | 1,000 | 2,320 |
| 17 | Kidney Renal Papillary Cell Carcinoma | KIRP | 292 | 0 | 100 | 392 |
| 18 | Brain Lower Grade Glioma | LGG | 594 | 0 | 1,000 | 1,594 |
| 19 | Liver Hepatocellular Carcinoma | LIHC | 377 | 0 | 200 | 577 |
| 20 | Lung Adenocarcinoma | LUAD | 585 | 1,139 | 2,000 | 3,724 |
| 21 | Lung Squamous Cell Carcinoma | LUSC | 539 | 1,081 | 500 | 2,120 |
| 22 | Mesothelioma | MESO | 87 | 0 | 200 | 287 |
| 23 | Ovarian Serous Cystadenocarcinoma | OV | 439 | 0 | 2,000 | 2,439 |
| 24 | Pancreatic Adenocarcinoma | PAAD | 185 | 557 | 3,000 | 3,742 |
| 25 | Pheochromocytoma and Paraganglioma | PCPG | 179 | 0 | 20 | 199 |
| 26 | Prostate Adenocarcinoma | PRAD | 498 | 0 | 3,000 | 3,498 |
| 27 | Rectum Adenocarcinoma | READ | 170 | 0 | 1,000 | 1,170 |
| 28 | Salivary Gland Cancer | SACA | 0 | 0 | 200 | 200 |
| 29 | Sarcoma (non-GIST) | SARC | 261 | 300 | 2,000 | 2,561 |
| 30 | Skin Cutaneous Melanoma | SKCM | 469 | 404 | 1,000 | 1,873 |
| 31 | Small Bowel Cancer | SBC | 0 | 0 | 50 | 50 |
| 32 | Stomach Adenocarcinoma | STAD | 443 | 0 | 1,000 | 1,443 |
| 33 | Testicular Germ Cell Tumors | TGCT | 150 | 0 | 500 | 650 |
| 34 | Thyroid Carcinoma | THCA | 507 | 0 | 1,000 | 1,507 |
| 35 | Thymoma | THYM | 124 | 0 | 20 | 144 |
| 36 | Uterine Corpus Endometrial Carcinoma | UCEC | 548 | 888 | 2,000 | 3,436 |
| 37 | Uterine Carcinosarcoma | UCS | 57 | 0 | 200 | 257 |
| 38 | Uterine Sarcoma | USARC | 0 | 0 | 500 | 500 |
| 39 | Uveal Melanoma | UVM | 80 | 0 | 50 | 130 |
| TOTAL | 10,438 | 6,004 | 34,690 | 51,132 |
🎯 Your Challenge: Build the Next Generation of Pathology Foundation Models
This competition is specifically designed for developing novel self-supervised learning algorithms and architectures to create powerful pathology foundation models.
What You’ll Develop:
- Self-supervised learning approaches that can learn meaningful representations from pathology images without requiring labeled data
- Novel neural network architectures optimized for the unique characteristics of histopathology data
- Foundation models that can serve as powerful feature extractors for diverse downstream clinical tasks
The Innovation Opportunity:
SLC-PFM focuses on pre-training methodologies - giving you the freedom to explore cutting-edge self-supervised techniques like:
- Contrastive Learning: SimCLR, MoCo, SwAV, VICReg
- Self-Distillation: DINO, DINOv2, BYOL
- Masked Image Modeling: MAE, iBOT, BEiT
- Multi-Scale Learning: Hierarchical representations across 20× and 40× magnifications
- Pathology-Specific Innovations: Novel augmentations, rotation invariance, cross-tile relationship modeling
- Transformer Architectures: Vision Transformers (ViTs) optimized for gigapixel pathology images
No labels. No supervision. Pure algorithmic innovation.
Design self-supervised learning frameworks that extract clinically meaningful patterns from 300M pathology images to create foundation models that will advance cancer diagnosis worldwide.
Technical Requirements
Development Environment
- Participants use their own computational infrastructure
- Self-supervised learning approaches without outcome labels
- Focus on developing generalizable pathology foundation models
Submission Requirements
By October 15, 2025, teams must provide:
- Foundation Model Checkpoint: Trained model weights
- Feature Extraction Code: Python implementation for inference
- Methodology Documentation: LaTeX-formatted technical description
Performance Requirements
- Submiited models must run feature extraction (inference) on a single GPU with ≤80GB memory
- Efficient feature extraction for downstream task evaluation
Data Distribution Options
- Secure link will be provided to access Phase 1 dataset after registration verification
Starting Kit (Available in June 2025 only to registered participants)
The comprehensive Starting Kit will include:
- Phase 1 Dataset: Complete MSK-SLCPFM collection
- Data Loading Scripts: Python code for loading, inspecting, and preprocessing WSI and tile data
- Documentation: Detailed guides on dataset structure and submission requirements
- Submission Resources: Templates for inference code (Python) and algorithm descriptions (LaTeX)
Unique Challenges of Pathology Data
The MSK-SLCPFM dataset presents unique challenges compared to natural image datasets:
Limited Color Variation
- H&E staining constrains colors to blue-purple (nuclei) and pink (cytoplasm/matrix)
- Different from rich color variation in natural images
Rotational Invariance
- Microscopic structures (nuclei, cells, glands) appear in arbitrary orientations
- No canonical orientation unlike objects in natural images
Non-IID Samples
- Tiles from same WSI are not completely independent samples
- Contrasts with ImageNet where images are Independent and Identically Distributed
Multi-Resolution Context
- Tissue details vary significantly across magnifications
- Glands appear larger with more cellular details at 40× vs 20× resolution
- Important pathological features may only be visible at specific magnifications
Data Ethics and Privacy
- All competition data has been de-identified
- Approved for research use under respective institutional IRB protocols
- Complies with HIPAA regulations
- Usage restricted to non-commercial academic research
- No patient overlap between Phase 1 and Phase 2 datasets
Questions about Phase 1? Contact: slcpfm2025@gmail.com