We find that three leading single-cell models—PIFiA, Paired Cell Inpainting, and DeepLoc—are sensitive to background context.
Classification accuracy drops by up to 15.8% when background cell localization differs from the center cell, and improves when they match.
