SlideCheck tool lets pathologists curate AI pretraining data with malignancy scores

Pathology foundation models are pretrained on massive streams of patches from whole-slide images (WSIs), but supervision is often sparse slide-level labels. This mismatch makes it hard to control which biological patterns enter the training data. SlideCheck addresses this by providing explicit abnormality and malignancy scores for each patch, using frozen features from an existing pathology foundation model. A dual-head MLP separately models broad abnormal morphology and malignant evidence, while a regularized feature-space scorer provides a supervised anchor. Score-attention agreement combines patch scores with WSI-level MIL attention to mine high-confidence pseudo-labels, enabling precise dataset curation.

Experiments show that SlideCheck-defined data distributions influence downstream self-supervised ViT pretraining, confirming biological composition as a controllable factor. Curated subsets achieve near full-data performance, meaning researchers can filter out irrelevant or low-quality patches without sacrificing model accuracy. This makes pretraining more efficient and auditable, turning large, undifferentiated patch pools into reusable datasets. The tool positions itself as a data guidance and auditing layer, helping labs build more trustworthy pathology AI with less data and clearer oversight.