SlideCheck: Guiding Self-Supervised Pretraining of Pathology Foundation Models via Dataset Distributions

Researchers introduce SlideCheck, a data guidance tool for pathology foundation models that uses frozen model features to score and curate pretraining datasets. The system provides abnormality and malignancy scores to help organize and audit WSI-derived patch data, demonstrating that controlled dataset composition significantly influences downstream self-supervised learning outcomes.

SlideCheck addresses a critical infrastructure challenge in pathology AI development: the disconnect between patch-level pretraining data and slide-level supervision signals. Pathology foundation models typically train on massive streams of image patches derived from whole-slide images (WSIs), yet labeling often occurs at coarser granularities with inconsistent quality. This mismatch obscures which biological patterns dominate pretraining, making models difficult to control and audit.

The tool operates as a data curation layer rather than a diagnostic model itself. Using a dual-head MLP architecture, SlideCheck independently estimates abnormality and malignancy signals for individual patches. By anchoring these estimates to supervised signals through a regularized feature-space scorer and leveraging slide-level multiple instance learning attention, the system generates high-confidence pseudo-labels that organize the underlying data. This enables construction of targeted pretraining subsets where pathologists and engineers retain explicit control over biological composition.

The research demonstrates that pretraining subset selection measurably affects downstream model behavior, validating dataset composition as an engineerable variable in foundation model development. Results indicating that curated subsets approach full-dataset performance suggest potential efficiency gains and reduced computational waste during pretraining. For the broader pathology AI ecosystem, SlideCheck represents a shift toward transparent, auditable data practices. This approach enables reproducibility and trust in clinical-grade AI systems where stakeholders increasingly demand visibility into training data construction. The framework likely influences how organizations building pathology foundation models approach dataset governance and quality assurance protocols.

• →SlideCheck provides explicit patch-level abnormality and malignancy scoring to guide curation of pathology pretraining datasets.
• →Dataset composition directly influences downstream behavior of self-supervised pathology models, making it an important controllable design factor.
• →Curated subsets can achieve comparable performance to full datasets, suggesting opportunities for more efficient and auditable pretraining.
• →The tool functions as a data guidance and auditing layer built on frozen foundation model features without requiring separate diagnostic capabilities.
• →Score-attention agreement mechanism mines high-confidence pseudo-labels by combining patch scores with WSI-level multiple instance learning signals.