Aggregating pretrained UNI2-h features extracted from the TCGA, CPTAC, and PANDA pathology image collections. The data consists of embeddings derived from 256 x 256 pixel patches processed at 20x magnification.
Use Cases
- Train slide-level classification models using the UNI2-h feature vectors from the TCGA cohort
- Perform cross-dataset validation for pathology tasks using features from CPTAC and PANDA
- Implement attention-based pooling on the 256x256 patch features to predict diagnostic labels
Strengths
- Features extracted from three major pathology datasets: TCGA, CPTAC, and PANDA
- Standardized patch dimensions of 256 x 256 pixels
- Consistent 20x magnification across all processed image patches
- Pre-computed embeddings generated via the UNI2-h model architecture