Decomposing Data Scale via Sample Count, Source, and Spatial Coverage

This study decomposes data scaling into unique sample count, source diversity, and spatial coverage using 11.6 million histology image patches. Results show that for spatially structured data, performance scales with broader spatial coverage and source diversity even at a fixed sample count.