Can single microscopy images match multi-channel quality using teacher models?

Medical microscopy typically needs multiple fluorescent channels (nuclear, membrane) to segment tissue accurately, but that's expensive and limits deployment. This work uses a frozen foundation model (SAM ViT-H) trained on full multi-channel images to teach a tiny student network (1.5–27M parameters) using only the nuclear channel. The distillation combines probability matching, boundary-aware losses, and uncertainty weighting. On TissueNet, the SAM-distilled Swin-Tiny student jumps from 65.31 to 78.36 Dice—recovering 88% of the full-channel oracle's performance—consistently improving all four tested architectures by ~12 points without retraining across datasets.